Microsound

This page intentionally left blank

Curtis Roads

Microsound

The MIT Press

Cambridge, Massachusetts
London, England

( 2001 Massachusetts Institute of Technology

All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage
and retrieval) without permission in writing from the publisher.
This book was set in Times New Roman in `3B2' by Asco Typesetters, Hong Kong and
was printed and bound in the United States of America.
Library of Congress Cataloging-in-Publication Data
Roads, Curtis.
Microsound / Curtis Roads.
p. cm.
Includes bibliographical references and index.
ISBN 0-262-18215-7 (hc. : alk. paper)
1. MusicAcoustics and physics. 2. Electronic musicHistory and criticism.
3. Computer musicHistory and criticism. I. Title.
ML3805 .R69 2001
2001030633
781.20 2dc21

Contents

Introduction

vii

Acknowledgments
Overview

ix

xi

Time Scales of Music

The History of Microsound from Antiquity to the Analog Era

Granular Synthesis

Varieties of Particle Synthesis

119

Transformation of Microsound

179

Windowed Analysis and Transformation

Microsound in Composition

Aesthetics of Composing with Microsound

Conclusion

References

85

301

349

353

Appendixes
A

The Cloud Generator Program

Sound Examples on the CD

Name Index
Subject Index

399
403

235

383
389

325

43

This page intentionally left blank

Introduction

Beneath the level of the note lies the realm of microsound, of sound particles. Microsonic particles remained invisible for centuries. Recent technological
advances let us probe and explore the beauties of this formerly unseen world.
Microsonic techniques dissolve the rigid bricks of music architecturethe notes
into a more uid and supple medium. Sounds may coalesce, evaporate, or
mutate into other sounds.
The sensations of point, pulse (regular series of points), line (tone), and surface (texture) appear as the density of particles increases. Sparse emissions leave
rhythmic traces. When the particles line up in rapid succession, they induce the
illusion of tone continuity that we call pitch. As the particles meander, they
ow into streams and rivulets. Dense agglomerations of particles form swirling
sound clouds whose shapes evolve over time.
In the 1940s, the Nobel prize winning physicist Dennis Gabor proposed that
any sound could be decomposed into acoustical quanta bounded by discrete
units of time and frequency. This quantum representation formed the famous
Gabor matrix. Like a sonogram, the vertical dimension of the Gabor matrix
indicated the location of the frequency energy, while the horizontal dimension
indicated the time region in which this energy occurred. In a related project,
Gabor built a machine to granulate sound into particles. This machine could
alter the duration of a sound without shifting its pitch.
In these two projects, the matrix and the granulator, Gabor accounted for
both important domains of sound representation. The matrix was the original
windowed frequency-domain representation. ``Windowed'' means segmented in
time, and ``frequency-domain'' refers to spectrum. The granulation machine, on
the other hand, operated on a time-domain representation, which is familiar to
anyone who has seen waveforms in a sound editor. This book explores microsound from both perspectives: the windowed frequency-domain and the micro

viii

Introduction

time-domain. Both concern microacoustic phenomena lasting less than onetenth of a second.
This book is the fruit of a lengthy period of activity involving synthesis
experiments, programming, and composition dating back to the early 1970s.
I started writing the text in 1995, after completing my textbook The Computer
Music Tutorial (The MIT Press 1996). Beginning with a few strands, it eventually grew into a lattice of composition theory, historical accounts, technical
overviews, acoustical experiments, descriptions of musical works, and aesthetic
reections. Why such a broad approach? Because at this stage of development,
the musical, technical, and aesthetic problems interweave. We are inventing
particles at the same time that we are learning how to compose with them. In
numerous ``assessment'' sections I have tried to summarize the results, which in
some cases are merely preliminary. More experimentation is surely needed.
Microsound records this rst round of experimentation, and thus serves as a
diary of research. Certain details, such as the specic software and hardware
that I used, will no doubt become obsolete rapidly. Even so, I decided to leave
them in for the historical record.
The experimentation and documentation could go on indenitely. One could
imagine, for example, a kind of synthesis ``cookbook'' after the excellent example of Jean-Claude Risset (1969). His text provided detailed recipes for making
specic sounds from a variety of synthesis techniques. This would be a worthy
project, and I would encourage others in this direction. As for myself, it is time
to compose.

Acknowledgments

This book derives from a doctoral thesis written for the Universite de Paris VIII
(Roads 1999). It would never have started without strong encouragement from
Professor Horacio Vaggione. I am deeply indebted to him for his patient advocacy, as well as for his inspired writings and pieces.
The congenial atmosphere in the Departement Musique at the Universite
de Paris VIII was ideal for the gestation of this work. I would also like to extend my sincere appreciation to Jean-Claude Risset and Daniel Arb. Despite
much pressure on their time, these pioneers and experts kindly agreed to
serve on the doctoral committee. Their commentaries on my text resulted in
major improvements.
I owe a debt of thanks to my colleague Gerard Pape at the Centre de
Creation Musicale Iannis Xenakis (CCMIX) for his support of my research,
teaching, and composition. I must also convey appreciation to Iannis Xenakis
for his brilliant example and for his support of our work in Paris. My rst
contact with him, at his short course in Formalized Music in 1972, started me
on this path.
I completed this book while teaching in the Center for Research in Electronic
Art Technology (CREATE) in the Department of Music and in the Media
Arts and Technology Program at the University of California, Santa Barbara.
I greatly appreciate the friendship and support of Professor JoAnn KucheraMorin, Director of CREATE, during this productive period. I would also like
to extend my thanks to the rest of the CREATE team, including Stephen T.
Pope for his collaboration on pulsar synthesis in 1997. It was a great pleasure to
work with Alberto de Campo, who served as CREATE's Research Director
in 19992000. Together we developed the PulsarGenerator software and the
Creatovox synthesizer. I consider these engaging musical instruments to be
among the main accomplishments of this research.

Acknowledgments

Allow me to remember my late colleague Professor Aldo Piccialli of the

Department of Physics at the University of Naples Federico II. His intense
fascination with the subject of microsound inspired me to dive deeper into the
theory of musical signal processing, and led to the notion of pulsar synthesis.
This exploration has been most rewarding. I also appreciate the discussions
and correspondence with my friends in Naples, including Sergio Cavaliere, Gianpaolo Evangelista, and Giancarlo Sica.
Many other colleagues provided information and advice, including Clifton
Kussmaul, Corey Cheng, Tom Erbe, Christopher Weare, and Jean de Reydellet. Brigitte Robindore, Pierre Roy, Jakub Omsky, Luca Lucchese, and Thom
Blum kindly read parts of the manuscript and provided much valuable feedback. Their comments are most appreciated. The MIT Press arranged for three
anonymous reviews of the book. These critiques led to many improvements. I
would also like to thank Douglas Sery of The MIT Press for his enthusiastic
sponsorship of this project.
Parts of this book were written during vacations at the family home in Illinois. I will always be grateful to my mother, Marjorie Roads, for the warm
atmosphere that I enjoyed there during sabbaticals.

Overview

Chapter 1 projects a view of nine time scales of musical sound structure. It

examines this hierarchy from both aesthetic and technical viewpoints. Major
themes of this chapter include: the boundaries between time scales, the particularities of the various time scales, and the size of sounds.
Chapter 2 traces the history of the idea of microsound, from the ancient
philosophy of atomism to the recent analog era. It explains how particle models
of sound emerged alongside wave-oriented models. It then presents the modern
history of microsound, beginning with the Gabor matrix. It follows the writings
of a diverse collection of authors, including Ezra Pound, Henry Cowell, Werner
Meyer-Eppler, Iannis Xenakis, Abraham Moles, Norbert Wiener, and Karlheinz Stockhausen. It also looks at the viability of a microsonic approach in
analog synthesis and instrumental music.
Chapter 3 presents the theory and practice of digital granular synthesis in its
myriad manifestations. It examines the dierent methods for organizing the
grains, and looks at the eects produced in each parameter of the technique. It
then surveys the various implementations of computer-based granular synthesis, beginning with the earliest experiments in the 1970s.
Chapter 4 is a catalog of experiments with newer particles, featuring glissons,
grainlets, pulsars, and trainlets. We also examine sonographic and formant
particles, transient drawing, particle cloning, and physical and abstract models
of particle synthesis.
Chapter 5 surveys a broad variety of microsonic sound transformations.
These range from audio compression techniques to micromontage and granulations. The brief presentation on the Creatovox instrument emphasizes real-time
performance with granulated sound. The chapter then covers transformations
on a micro scale, including pitch-shifting, pitch-time changing, ltering, dynamics processing, frequency-domain granulation, and waveset transformations.

xii

Overview

The nal sections present techniques of spatialization with sound particles, and
convolution with microsounds.
Chapter 6 explores a variety of sound transformations based on windowed
spectrum analysis. After a theoretical section, it presents the main tools of windowed spectrum transformation, including the phase vocoder, the tracking
phase vocoder, the wavelet transform, and Gabor analysis.
Chapter 7 turns from technology to compositional applications. It begins
with a description of the rst studies realized with granular synthesis on a digital computer. It then looks at particle techniques in my recent compositions, as
well as those by Barry Truax, Horacio Vaggione, and other composers.
Chapter 8, on the aesthetics of composing with microsound, is the most
philosophical part of the book. It highlights both specic and general aesthetic
issues raised by microsound in composition.
Chapter 9 concludes with a commentary on the future of microsound in
music.

Microsound

Time Scales of Music

Time Scales of Music

Boundaries between Time Scales
Zones of Intensity and Frequency
Innite Time Scale
Supra Time Scale
Macro Time Scale
Perception of the Macro Time Scale
Macroform
Design of Macroform
Meso Time Scale
Sound Masses, Textures, and Clouds
Cloud Taxonomy
Sound Object Time Scale
The Sensation of Tone
Homogeneous Notes versus Heterogeneous Sound Objects
Sound Object Morphology
Micro Time Scale
Perception of Microsound
Microtemporal Intensity Perception
Microtemporal Fusion and Fission

Chapter 1

Microtemporal Silence Perception

Microtemporal Pitch Perception
Microtemporal Auditory Acuity
Microtemporal Preattentive Perception
Microtemporal Subliminal Perception
Viewing and Manipulating the Microtime Level
Do the Particles Really Exist?
Heterogeneity in Sound Particles
Sampled Time Scale
Sound Composition with Individual Sample Points
Assessment of Sound Composition with Samples
Subsample Time Scale
Aliased Artefacts
Ultrasonic Loudspeakers
Atomic Sound: Phonons and Polarons
At the Physical Limits: the Planck Time Interval
Innitesimal Time Scale
Outside Time Music
The Size of Sounds
Summary
The evolution of musical expression intertwines with the development of musical instruments. This was never more evident than in the twentieth century.
Beginning with the gigantic Telharmonium synthesizer unveiled in 1906 (Weidenaar 1989, 1995), research ushered forth a steady stream of electrical and
electronic instruments. These have irrevocably molded the musical landscape.
The most precise and exible electronic music instrument ever conceived is
the digital computer. As with the pipe organ, invented centuries earlier, the
computer's power derives from its ability to emulate, or in scientic terms, to
model phenomena. The models of the computer take the form of symbolic
code. Thus it does not matter whether the phenomena being modeled exist
outside the circuitry of the machine, or whether they are pure fantasy. This

Time Scales of Music

makes the computer an ideal testbed for the representation of musical structure
on multiple time scales.
This chapter examines the time scales of music. Our main focus is the micro
time scale and its interactions with other time scales. By including extreme time
scalesthe innite and the innitesimalwe situate musical time within the
broadest possible context.

Time Scales of Music

Music theory has long recognized a temporal hierarchy of structure in music
compositions. A central task of composition has always been the management
of the interaction amongst structures on dierent time scales. Starting from the
topmost layer and descending, one can dissect layers of structure, arriving at
the bottom layer of individual notes.
This hierarchy, however, is incomplete. Above the level of an individual piece
are the cultural time spans dening the oeuvre of a composer or a stylistic
period. Beneath the level of the note lies another multilayered stratum, the
microsonic hierarchy. Like the quantum world of quarks, leptons, gluons,
and bosons, the microsonic hierarchy was long invisible. Modern tools let us
view and manipulate the microsonic layers from which all acoustic phenomena
emerge. Beyond these physical time scales, mathematics denes two ideal
temporal boundariesthe innite and the innitesimalwhich appear in the
theory of musical signal processing.
Taking a comprehensive view, we distinguish nine time scales of music,
starting from the longest:
1. Innite The ideal time span of mathematical durations such as the innite
sine waves of classical Fourier analysis.
2. Supra A time scale beyond that of an individual composition and extending into months, years, decades, and centuries.
3. Macro The time scale of overall musical architecture or form, measured in
minutes or hours, or in extreme cases, days.
4. Meso Divisions of form. Groupings of sound objects into hierarchies of
phrase structures of various sizes, measured in minutes or seconds.
5. Sound object A basic unit of musical structure, generalizing the traditional
concept of note to include complex and mutating sound events on a time
scale ranging from a fraction of a second to several seconds.

Chapter 1

6. Micro Sound particles on a time scale that extends down to the threshold of auditory perception (measured in thousandths of a second or milliseconds).
7. Sample The atomic level of digital audio systems: individual binary samples or numerical amplitude values, one following another at a xed time
interval. The period between samples is measured in millionths of a second
(microseconds).
8. Subsample Fluctuations on a time scale too brief to be properly recorded
or perceived, measured in billionths of a second (nanoseconds) or less.
9. Innitesimal The ideal time span of mathematical durations such as the
innitely brief delta functions.
Figure 1.1 portrays the nine time scales of the time domain. Notice in the
middle of the diagram, in the frequency column, a line indicating ``Conscious
time, the present (@600 ms).'' This line marks o Winckel's (1967) estimate of
the ``thickness of the present.'' The thickness extends to the line at the right
indicating the physical NOW. This temporal interval constitutes an estimate
of the accumulated lag time of the perceptual and cognitive mechanisms associated with hearing. Here is but one example of a disparity between chronos
physical time, and tempusperceived time (Kupper 2000).
The rest of this chapter explains the characteristics of each time scale in turn.
We will, of course, pay particular attention to the micro time scale.

Boundaries between Time Scales

As sound passes from one time scale to another it crosses perceptual boundaries. It seems to change quality. This is because human perception processes
each time scale dierently. Consider a simple sinusoid transposed to various
time scales (1 msec, 1 ms, 1 sec, 1 minute, 1 hour). The waveform is identical,
but one would have diculty classifying these auditory experiences in the same
family.
In some cases the borders between time scales are demarcated clearly; ambiguous zones surround others. Training and culture condition perception of
the time scales. To hear a at pitch or a dragging beat, for example, is to
detect a temporal anomaly on a micro scale that might not be noticed by other
people.

Time Scales of Music

Figure 1.1 The time domain, segmented into periods, time delay eects, frequencies,
and perception and action. Note that time intervals are not drawn to scale.

Chapter 1

Digital audio systems, such as compact disc players, operate at a xed sampling frequency. This makes it easy to distinguish the exact boundary separating the sample time scale from the subsample time scale. This boundary is the
Nyquist frequency, or the sampling frequency divided by two. The eect of
crossing this boundary is not always perceptible. In noisy sounds, aliased frequencies from the subsample time domain may mix unobtrusively with high
frequencies in the sample time domain.
The border between certain other time scales is context-dependent. Between
the sample and micro time scales, for example, is a region of transient events
too brief to evoke a sense of pitch but rich in timbral content. Between the
micro and the object time scales is a stratum of brief events such as short staccato notes. Another zone of ambiguity is the border between the sound object
and meso levels, exemplied by an evolving texture. A texture might contain a
statistical distribution of micro events that are perceived as a unitary yet timevarying sound.
Time scales interlink. A given level encapsulates events on lower levels and is
itself subsumed within higher time scales. Hence to operate on one level is to
aect other levels. The interaction between time scales is not, however, a simple
relation. Linear changes on a given time scale do not guarantee a perceptible
eect on neighboring time scales.

Zones of Intensity and Frequency

Sound is an alternation in pressure, particle displacement, or particle velocity propagated
in an elastic material. (Olson 1957)

Before we continue further, a brief discussion of acoustic terminology might be

helpful. In scientic parlanceas opposed to popular usagethe word ``sound''
refers not only to phenomena in air responsible for the sensation of hearing
but also ``whatever else is governed by analogous physical principles'' (Pierce
1994). Sound can be dened in a general sense as mechanical radiant energy
that is transmitted by pressure waves in a material medium. Thus besides the
airborne frequencies that our ears perceive, one may also speak of underwater
sound, sound in solids, or structure-borne sound. Mechanical vibrations even
take place on the atomic level, resulting in quantum units of sound energy called
phonons. The term ``acoustics'' likewise is independent of air and of human
perception. It is distinguished from optics in that it involves mechanicalrather
than electromagnetic, wave motion.

Time Scales of Music

Corresponding to this broad denition of sound is a very wide range of

transient, chaotic, and periodic uctuations, spanning frequencies that are both
higher and lower than the human ear can perceive. The audio frequencies, traditionally said to span the range of about 20 Hz to 20 kHz are perceptible to the
ear. The specic boundaries vary depending on the individual.
Vibrations at frequencies too low to be heard as continuous tones can be
perceived by the ear as well as the body. These are the infrasonic impulses and
vibrations, in the range below about 20 Hz. The infectious rhythms of the percussion instruments fall within this range.
Ultrasound includes the domain of high frequencies above the range of human audibility. The threshold of ultrasound varies according to the individual,
their age, and the test conditions. Science and industry use ultrasonic techniques in a variety of applications, such as acoustic imaging (Quate 1998) and
highly directional loudspeakers (Pompei 1998).
Some sounds are too soft to be perceived by the human ear, such as a caterpillar's delicate march across a leaf. This is the zone of subsonic intensities.
Other sounds are so loud that to perceive them directly is dangerous, since
they are destructive to the human body. Sustained exposure to sound levels
around 120 dB leads directly to pain and hearing loss. Above 130 dB, sound is
felt by the exposed tissues of the body as a painful pressure wave (Pierce 1983).
This dangerous zone extends to a range of destructive acoustic phenomena. The
force of an explosion, for example, is an intense acoustic shock wave.
For lack of a better term, we call these perisonic intensities (from the Latin
periculos meaning ``dangerous''). The audible intensities fall between these two
ranges. Figure 1.2 depicts the zones of sound intensity and frequency. The a
zone in the center is where audio frequencies intersect with audible intensities,
enabling hearing. Notice that the a zone is but a tiny fraction of a vast range of
sonic phenomena.
Following this discussion of acoustical terms, let us proceed to the main
theme of this chapter, the time scales of music.

Innite Time Scale

Complex Fourier analysis regards the signal sub specie aeternitatis.

(Gabor 1952)

The human experience of musical time is linked to the ticking clock. It is

natural to ask: when did the clock begin to tick? Will it tick forever? At the

Chapter 1

Figure 1.2 Zones of intensities and frequencies. Only the zone marked a is audible to
the ear. This zone constitutes a tiny portion of the range of sound phenomena.

extreme upper boundary of all time scales is the mathematical concept of an

innite time span. This is a logical extension of the innite series, a fundamental
notion in mathematics. An innite series is a sequence of numbers u1 ; u 2 ; u 3 . . .
arranged in a prescribed order and formed according to a particular rule.
Consider this innite series:
y
X

ui u1 u 2 u 3

i1

This equation sums a set of numbers ui , where the index i goes from 1 to y.
What if each number ui corresponded to a tick of a clock? This series would
then dene an innite duration. This ideal is not so far removed from music as
it may seem. The idea of innite duration is implicit in the theory of Fourier
analysis, which links the notion of frequency to sine waves of innite duration.
As chapter 6 shows, Fourier analysis has proven to be a useful tool in the analysis and transformation of musical sound.

Time Scales of Music

Figure 1.3

The scope of the supratemporal domain.

Supra Time Scale

The supra time scale spans the durations that are beyond those of an individual
composition. It begins as the applause dies out after the longest compositions, and extends into weeks, months, years, decades, and beyond (gure 1.3).
Concerts and festivals fall into this category. So do programs from music
broadcasting stations, which may extend into years of more-or-less continuous
emissions.
Musical cultures are constructed out of supratemporal bricks: the eras of
instruments, of styles, of musicians, and of composers. Musical education takes
years; cultural tastes evolve over decades. The perception and appreciation of

10

Chapter 1

a single composition may change several times within a century. The entire
history of music transpires within the supratemporal scale, starting from the
earliest known musical instrument, a Neanderthal ute dating back some
45,000 years (Whitehouse 1999).
Composition is itself a supratemporal activity. Its results last only a fraction
of the time required for its creation. A composer may spend a year to complete
a ten-minute piece. Even if the composer does not work every hour of every
day, the ratio of 52,560 minutes passed for every 1 minute composed is still
signicant. What happens in this time? Certain composers design a complex
strategy as prelude to the realization of a piece. The electronic music composer
may spend considerable time in creating the sound materials of the work. Either
of these tasks may entail the development of software. Virtually all composers
spend time experimenting, playing with material in dierent combinations.
Some of these experiments may result in fragments that are edited or discarded,
to be replaced with new fragments. Thus it is inevitable that composers invest
time pursuing dead ends, composing fragments that no one else will hear. This
backtracking is not necessarily time wasted; it is part of an important feedback
loop in which the composer renes the work. Finally we should mention documentation. While only a few composers document their labor, these documents
may be valuable to those seeking a deeper understanding of a work and the
compositional process that created it. Compare all this with the eciency of the
real-time improviser!
Some music spans beyond the lifetime of the individual who composed it,
through published notation, recordings, and pedagogy. Yet the temporal reach
of music is limited. Many compositions are performed only once. Scores, tapes,
and discs disappear into storage, to be discarded sooner or later. Music-making
presumably has always been part of the experience of Homo sapiens, who it is
speculated came into being some 200,000 years ago. Few traces remain of
anything musical older than a dozen centuries. Modern electronic instruments
and recording media, too, are ephemeral. Will human musical vibrations somehow outlast the species that created them? Perhaps the last trace of human
existence will be radio waves beamed into space, traveling vast distances before
they dissolve into noise.
The upper boundary of time, as the concept is currently understood, is the
age of the physical universe. Some scientists estimate it to be approximately
fteen billion years (Lederman and Scramm 1995). Cosmologists continue to
debate how long the universe may expand. The latest scientic theories continue to twist the notion of time itself (see, for example, Kaku 1995; ArkaniHamed et al. 2000).

11

Time Scales of Music

Macro Time Scale

The macro level of musical time corresponds to the notion of form, and
encompasses the overall architecture of a composition. It is generally measured
in minutes. The upper limit of this time scale is exemplied by such marathon
compositions as Richard Wagner's Ring cycle, the Japanese Kabuki theater,
Jean-Claude Eloy's evening-long rituals, and Karlheinz Stockhausen's opera
Licht (spanning seven days and nights). The literature of opera and contemporary music contains many examples of music on a time scale that exceeds two
hours. Nonetheless, the vast majority of music compositions realized in the past
century are less than a half-hour in duration. The average duration is probably
in the range of a kilosecond (16 min 40 sec). Complete compositions lasting less
than a hectosecond (1 min 40 sec) are rare.
Perception of the Macro Time Scale
Unless the musical form is described in advance of performance (through program notes, for example), listeners perceive the macro time scale in retrospect,
through recollection. It is common knowledge that the remembrance of things
past is subject to strong discontinuities and distortions. We cannot recall time
as a linearly measured ow. As in everyday life, the perceived ow of musical
time is linked to reference events or memories that are tagged with emotional
signicance.
Classical music (Bach, Mozart, Beethoven, etc.) places reference events at
regular intervals (cadences, repetition) to periodically orient the listener within
the framework of the form. Some popular music takes this to an extreme,
reminding listeners repeatedly on a shorter time base.
Subjective factors play into a distorted sense of time. Was the listener engaged in aesthetic appreciation of the work? Were they paying attention? What
is their musical taste, their training? Were they preoccupied with stress and
personal problems? A composition that we do not understand or like appears
to expand in time as we experience it, yet vanishes almost immediately from
memory.
The perception of time ow also depends on the objective nature of the musical materials. Repetition and a regular pulse tend to carry a work eciently
through time, while an unchanging, unbroken sound (or silence) reduces the
ow to a crawl.

12

Chapter 1

The ear's sensitivity to sound is limited in duration. Long continuous noises

or regular sounds in the environment tend to disappear from consciousness and
are noticed again only when they change abruptly or terminate.
Macroform
Just as musical time can be viewed in terms of a hierarchy of time scales, so it
is possible to imagine musical structure as a tree in the mathematical sense.
Mathematical trees are inverted, that is, the uppermost level is the root symbol,
representing the entire work. The root branches into a layer of macrostructure
encapsulating the major parts of the piece. This second level is the form: the
arrangement of the major sections of the piece. Below the level of form is a
syntactic hierarchy of branches representing mesostructures that expand into
the terminal level of sound objects (Roads 1985d).
To parse a mathematical tree is straightforward. Yet one cannot parse a sophisticated musical composition as easily as a compiler parses a computer program. A compiler references an unambiguous formal grammar. By contrast, the
grammar of music is ambiguoussubject to interpretation, and in a perpetual
state of evolution. Compositions may contain overlapping elements (on various
levels) that cannot be easily segmented. The musical hierarchy is often fractured. Indeed, this is an essential ingredient of its fascination.
Design of Macroform
The design of macroform takes one of two contrasting paths: top-down or
bottom-up. A strict top-down approach considers macrostructure as a preconceived global plan or template whose details are lled in by later stages of composition. This corresponds to the traditional notion of form in classical music,
wherein certain formal schemes have been used by composers as molds (Apel
1972). Music theory textbooks catalog the generic classical forms (Leichtentritt
1951) whose habitual use was called into question at the turn of the twentieth
century. Claude Debussy, for example, discarded what he called ``administrative forms'' and replaced them with uctuating mesostructures through a chain
of associated variations. Since Debussy, composers have written a tremendous
amount of music not based on classical forms. This music is full of local detail
and eschews formal repetition. Such structures resist classication within the
catalog of standard textbook forms. Thus while musical form has continued to
evolve in practice in the past century, the acknowledged catalog of generic forms
has hardly changed.

13

Time Scales of Music

This is not to say that the use of preconceived forms has died away. The
practice of top-down planning remains common in contemporary composition.
Many composers predetermine the macrostructure of their pieces according to
a more-or-less formal scheme before a single sound is composed.
By contrast, a strict bottom-up approach conceives of form as the result of a
process of internal development provoked by interactions on lower levels of
musical structure. This approach was articulated by Edgard Varese (1971), who
said, ``Form is a resultthe result of a process.'' In this view, macrostructure
articulates processes of attraction and repulsion (for example, in the rhythmic
and harmonic domains) unfolding on lower levels of structure.
Manuals on traditional composition oer myriad ways to project low-level
structures into macrostructure:
Smaller forms may be expanded by means of external repetitions, sequences, extensions,
liquidations and broadening of connectives. The number of parts may be increased by supplying codettas, episodes, etc. In such situations, derivatives of the basic motive are formulated into new thematic units. (Schoenberg 1967)

Serial or germ-cell approaches to composition expand a series or a formula

through permutation and combination into larger structures.
In the domain of computer music, a frequent technique for elaboration is to
time-expand a sound fragment into an evolving sound mass. Here the unfolding
of sonic microstructure rises to the temporal level of a harmonic progression.
A dierent bottom-up approach appears in the work of the conceptual and
chance composers, following in the wake of John Cage. Cage (1973) often conceived of form as arising from a series of accidentsrandom or improvised
events occurring on the sound object level. For Cage, form (and indeed sound)
was a side-eect of a conceptual strategy. Such an approach often results in
discontinuous changes in sound structure. This was not accidental; Cage disdained continuity in musical structure, always favoring juxtaposition:
Where people had felt the necessity to stick sounds together to make a continuity, we felt
the necessity to get rid of the glue so that sounds would be themselves. (Cage 1959)

For some, composition involves a mediation between the top-down and

bottom-up approaches, between an abstract high-level conception and the concrete materials being developed on lower levels of musical time structure. This
implies negotiation between a desire for orderly macrostructure and imperatives that emerge from the source material. Certain phrase structures cannot
be encapsulated neatly within the box of a precut form. They mandate a container that conforms to their shape and weight.

14

Chapter 1

The debate over the emergence of form is ancient. Musicologists have

long argued whether, for example, a fugue is a template (form) or a process
of variation. This debate echoes an ancient philosophical discourse pitting
form against ux, dating back as far as the Greek philosopher Heraclitus. Ultimately, the dichotomy between form and process is an illusion, a failure of
language to bind two aspects of the same concept into a unit. In computer
science, the concept of constraints does away with this dichotomy (Sussman and
Steele 1981). A form is constructed according to a set of relationships. A set of
relationships implies a process of evaluation that results in a form.

Meso Time Scale

The mesostructural level groups sound objects into a quasi hierarchy of phrase
structures of durations measured in seconds. This local as opposed to global
time scale is extremely important in composition, for it is most often on the
meso level that the sequences, combinations, and transmutations that constitute
musical ideas unfold. Melodic, harmonic, and contrapuntal relations happen
here, as do processes such as theme and variations, and many types of development, progression, and juxtaposition. Local rhythmic and metric patterns,
too, unfold on this stratum.
Wishart (1994) called this level of structure the sequence. In the context of
electronic music, he identied two properties of sequences: the eld (the material, or set of elements used in the sequence), and the order. The eld serves as
a lexiconthe vocabulary of a piece of music. The order determines thematic
relationsthe grammar of a particular piece. As Wishart observed, the eld and
the order must be established quickly if they are to serve as the bearers of musical
code. In traditional music, they are largely predetermined by cultural norms.
In electronic music, the meso layer presents timbre melodies, simultaneities
(chord analogies), spatial interplay, and all manner of textural evolutions. Many
of these processes are described and classied in Denis Smalley's interesting
theory of spectromorphologya taxonomy of sound gesture shapes (Smalley
1986, 1997).
Sound Masses, Textures, and Clouds
To the sequences and combinations of traditional music, we must add another
principle of organization on the meso scale: the sound mass. Decades ago,

15

Time Scales of Music

Edgard Varese predicted that the sounds introduced by electronic instruments

would necessitate new organizing principles for mesostructure.
When new instruments will allow me to write music as I conceive it, taking the place of the
linear counterpoint, the movement of sound masses, or shifting planes, will be clearly perceived. When these sound masses collide the phenomena of penetration or repulsion will
seem to occur. (Varese 1962)

A trend toward shaping music through the global attributes of a sound mass
began in the 1950s. One type of sound mass is a cluster of sustained frequencies
that fuse into a solid block. In a certain style of sound mass composition,
musical development unfolds as individual lines are added to or removed from
this cluster. Gyorgy Ligeti's Volumina for organ (1962) is a masterpiece of this
style, and the composer has explored this approach in a number of other pieces,
including Atmospheres (1961) and Lux Aeterna (1966).
Particles make possible another type of sound mass: statistical clouds of
microevents (Xenakis 1960). Wishart (1994) ascribed two properties to cloud
textures. As with sequences, their eld is the set of elements used in the texture,
which may be constant or evolving. Their second property is density, which
stipulates the number of events within a given time period, from sparse scatterings to dense scintillations.
Cloud textures suggest a dierent approach to musical organization. In
contrast to the combinatorial sequences of traditional meso structure, clouds
encourage a process of statistical evolution. Within this evolution the composer can impose specic morphologies. Cloud evolutions can take place in the
domain of amplitude (crescendi/decrescendi), internal tempo (accelerando/
rallentando), density (increasing/decreasing), harmonicity (pitch/chord/cluster/
noise, etc.), and spectrum (high/mid/low, etc.).
Xenakis's tape compositions Concret PH (1958), Bohor I (1962), and Persepolis (1971) feature dense, monolithic clouds, as do many of his works for
traditional instruments. Stockhausen (1957) used statistical form-criteria as one
component of his early composition technique. Since the 1960s, particle
textures have appeared in numerous electroacoustic compositions, such as the
remarkable De natura sonorum (1975) of Bernard Parmegiani.
Varese spoke of the interpenetration of sound masses. The diaphanous nature of cloud structures makes this possible. A crossfade between two clouds
results in a smooth mutation. Mesostructural processes such as disintegration
and coalescence can be realized through manipulations of particle density (see
chapter 6). Density determines the transparency of the material. An increase in

16

Chapter 1

density lifts a cloud into the foreground, while a decrease causes evaporation,
dissolving a continuous sound band into a pointillist rhythm or vaporous background texture.
Cloud Taxonomy
To describe sound clouds precisely, we might refer to the taxonomy of cloud
shapes in the atmosphere:
Cumulus

well-dened cauliower-shaped cottony clouds

Stratocumulus
Stratus

a thin fragmented layer, often translucent

Nimbostratus
Cirrus

blurred by wind motion

a widespread gray or white sheet, opaque

isolated sheets that develop in laments or patches

In another realm, among the stars, outer space is lled with swirling clouds of
cosmic raw material called nebulae.
The cosmos, like the sky on a turbulent summer day, is lled with clouds of dierent sizes,
shapes, structures, and distances. Some are swelling cumulus, others light, wispy cirrusall
of them constantly changing colliding, forming, and evaporating. (Kaler 1997)

Pulled by immense gravitational elds or blown by cosmic shockwaves,

nebulae form in great variety: dark or glowing, amorphous or ring-shaped,
constantly evolving in morphology. These forms, too, have musical analogies.
Programs for sonographic synthesis (such as MetaSynth [Wenger and Spiegel
1999]), provide airbrush tools that let one spray sound particles on the timefrequency canvas. On the screen, the vertical dimension represents frequency,
and the horizontal dimension represents time. The images can be blurred, fragmented, or separated into sheets. Depending on their density, they may be
translucent or opaque. Displacement maps can warp the cloud into a circular
or spiral shape on the time-frequency canvas. (See chapter 6 on sonographic
transformation of sound.)

Sound Object Time Scale

The sound object time scale encompasses events of a duration associated with
the elementary unit of composition in scores: the note. A note usually lasts from
about 100 ms to several seconds, and is played by an instrument or sung by a

17

Time Scales of Music

vocalist. The concept of sound object extends this to allow any sound, from
any source. The term sound object comes from Pierre Schaeer, the pioneer of
musique concrete. To him, the pure objet sonore was a sound whose origin a
listener could not identify (Schaeer 1959, 1977, p. 95). We take a broader view
here. Any sound within stipulated temporal limits is a sound object. Xenakis
(1989) referred to this as the ``ministructural'' time scale.
The Sensation of Tone
The sensation of tonea sustained or continuous event of denite or indenite
pitchoccurs on the sound object time scale. The low-frequency boundary for
the sensation of a continuous soundas opposed to a uttering succession of
brief microsoundshas been estimated at anywhere from 8 Hz (Savart) to
about 30 Hz. (As reference, the deepest sound in a typical orchestra is the open
E of the contrabass at 41.25 Hz.) Helmholtz, the nineteenth century German
acoustician, investigated this lower boundary.
In the rst place it is necessary that the strength of the vibrations of the air for very low
tones should be extremely greater than for high tones. The increase in strength . . . is of
especial consequence in the deepest tones. . . . To discover the limit of the deepest tones it is
necessary not only to produce very violent agitations in the air but to give these a simple
pendular motion. (Helmholtz 1885)

Helmholtz observed that a sense of continuity takes hold between 24 to 28

Hz, but that the impression of a denite pitch does not take hold until 40 Hz.
Pitch and tone are not the same thing. Acousticians speak of complex tones
and unpitched tones. Any sound perceived as continuous is a tone. This can, for
example include noise.
Between the sensation of a continuous tone and the sensation of metered
rhythm stands a zone of ambiguity, an infrasonic frequency domain that is too
slow to form a continuous tone but too fast for rhythmic denition. Thus continuous tone is a possible quality, but not a necessary property, of a sound object. Consider a relatively dense cloud of sonic grains with short silent gaps on
the order of tens of milliseconds. Dozens of dierent sonic events occur per
second, each unique and separated by a brief intervals of zero amplitude, yet
such a cloud is perceived as a unitary eventa single sound object.
A sense of regular pulse and meter begins to occur from approximately 8 Hz
down to 0.12 Hz and below (Fraisse 1982). Not coincidentally, it is in this
rhythmically apprensible range that the most salient and expressive vibrato,
tremolo, and spatial panning eects occur.

18

Chapter 1

Homogeneous Notes versus Heterogeneous Sound Objects

The sound object time scale is the same as that of traditional notes. What distinguishes sound objects from notes? The note is the homogeneous brick of
conventional music architecture. Homogeneous means that every note can be
described by the same four properties:
1 pitch, generally one of twelve equal-tempered pitch classes
1 timbre, generally one of about twenty dierent instruments for a full orchestra, with two or three dierent attack types for each instrument
1 dynamic marking, generally one of about ten dierent relative levels
1 duration, generally between @100 ms (slightly less than a thirty-second note
at a tempo of 60 M.M.) to @8 seconds (for two tied whole notes)
These properties are static, guaranteeing that, in theory, a note in one
measure with a certain pitch, dynamic, and instrumental timbre is functionally
equivalent to a note in another measure with the same three properties. The
properties of a pair of notes can be compared on a side-by-side basis and a
distance or interval can be calculated. The notions of equivalence and distance
lead to the notion of invariants, or intervallic distances that are preserved across
transformations.
Limiting material to a static homogeneous set allows abstraction and eciency in musical language. It serves as the basis for operations such as
transposition, orchestration and reduction, the algebra of tonal harmony and
counterpoint, and the atonal and serial manipulations. In the past decade, the
MIDI protocol has extended this homogeneity into the domain of electronic
music through standardized note sequences that play on any synthesizer.
The merit of this homogeneous system is clear; highly elegant structures
having been built with standard materials inherited from centuries past. But
since the dawn of the twentieth century, a recurring aesthetic dream has been
the expansion beyond a xed set of homogeneous materials to a much larger
superset of heterogeneous musical materials.
What we have said about the limitations of the European note concept does
not necessarily apply to the musics of other cultures. Consider the shakuhachi
music of Japan, or contemporary practice emerging from the advanced developments of jazz.
Heterogeneity means that two objects may not share common properties.
Therefore their percept may be entirely dierent. Consider the following two
examples. Sound A is a brief event constructed by passing analog diode noise

19

Time Scales of Music

through a time-varying bandpass lter and applying an exponentially decaying

envelope to it. Sound B lasts eight seconds. It is constructed by granulating
in multiple channels several resonant low-pitched strokes on an African slit
drum, then reverberating the texture. Since the amplitudes and onset times of
the grains vary, this creates a jittering sound mass. To compare A and B is like
comparing apples and oranges. Their microstructures are dierent, and we can
only understand them through the properties that they do not have in common.
Thus instead of homogeneous notes, we speak of heterogeneous sound objects.
The notion of sound object generalizes the note concept in two ways:
1. It puts aside the restriction of a common set of properties in favor of a heterogeneous collection of properties. Some objects may not share common
properties with other objects. Certain sound objects may function as unique
singularities. Entire pieces may be constructed from nothing but such
singularities.
2. It discards the notion of static, time-invariant properties in favor of timevarying properties (Roads 1985b).
Objects that do not share common properties may be separated into diverse
classes. Each class will lend itself to dierent types of manipulation and musical
organization. Certain sounds layer well, nearly any mixture of elongated sine
waves with smooth envelopes for example. The same sounds organized in a
sequence, however, rather quickly become boring. Other sounds, such as isolated impulses, are most eective when sparsely scattered onto a neutral sound
canvas.
Transformations applied to objects in one class may not be eective in another class. For example, a time-stretching operation may work perfectly well
on a pipe organ tone, preserving its identity and aecting only its duration. The
same operation applied to the sound of burning embers will smear the crackling
transients into a nondescript electronic blur.
In traditional western music, the possibilities for transition within a note are
limited by the physical properties of the acoustic instrument as well as frozen by
theory and style. Unlike notes, the properties of a sound object are free to vary
over time. This opens up the possibility of complex sounds that can mutate
from one state to another within a single musical event. In the case of synthesized sounds, an object may be controlled by multiple time-varying envelopes
for pitch, amplitude, spatial position, and multiple determinants of timbre.
These variations may take place over time scales much longer than those associated with conventional notes.

20

Chapter 1

We can subdivide a sound object not only by its properties but also by its
temporal states. These states are composable using synthesis tools that operate
on the microtime scale. The micro states of a sound can also be decomposed
and rearranged with tools such as time granulators and analysis-resynthesis
software.
Sound Object Morphology
In music, as in other elds, the organization is conditioned by the material.
1977, p. 680)

(Schaeer

The desire to understand the enormous range of possible sound objects led
Pierre Schaeer to attempt to classify them, beginning in the early 1950s
(Schaeer and Moles 1952). Book V of his Traite des objets musicaux (1977),
entitled Morphologie and typologie des objets sonores introduces the useful notion of sound object morphologythe comparison of the shape and evolution
of sound objects. Schaeer borrowed the term morphology from the sciences,
where it refers to the study of form and structure (of organisms in biology, of
word-elements in linguistics, of rocks in geology, etc.). Schaeer diagrammed
sound shape in three dimensions: the harmonic (spectrum), dynamic (amplitude), and melodic (pitch). He observed that the elements making up a complex sound can be perceived as either merged to form a sound compound, or
remaining separate to form a sound mixture. His typology, or classication
of sound objects into dierent groups, was based on acoustic morphological
studies.
The idea of sound morphology remains central to the theory of electroacoustic music (Bayle 1993), in which the musical spotlight is often shone on
the sound object level. In traditional composition, transitions function on the
mesostructural level through the interplay of notes. In electroacoustic music,
the morphology of an individual sound may play a structural role, and transitions can occur within an individual sound object. This ubiquity of mutation
means that every sonic event is itself a potential transformation.

Micro Time Scale

The micro time scale is the main subject of this book. It embraces transient
audio phenomena, a broad class of sounds that extends from the threshold of

21

Time Scales of Music

timbre perception (several hundred microseconds) up to the duration of short

sound objects (@100 ms). It spans the boundary between the audio frequency
range (approximately 20 Hz to 20 kHz) and the infrasonic frequency range
(below 20 Hz). Neglected in the past owing to its inaccessibility, the microtime
domain now stands at the forefront of compositional interest.
Microsound is ubiquitous in the natural world. Transient events unfold all
around in the wild: a bird chirps, a twig breaks, a leaf crinkles. We may not
take notice of microacoustical events until they occur en masse, triggering a
global statistical percept. We experience the interactions of microsounds in the
sound of a spray of water droplets on a rocky shore, the gurgling of a brook,
the pitter-patter of rain, the crunching of gravel being walked upon, the snapping of burning embers, the humming of a swarm of bees, the hissing of
rice grains poured into a bowl, and the crackling of ice melting. Recordings
of dolphins reveal a language made up entirely of high-frequency clicking
patterns.
One could explore the microsonic resources of any musical instrument in its
momentary bursts and infrasonic utterings, (a study of traditional instruments
from this perspective has yet to be undertaken). Among unpitched percussion,
we nd microsounds in the angled rainstick, (shaken) small bells, (grinding)
ratchet, (scraped) guiro, ( jingling) tambourine, and the many varieties of
rattles. Of course, the percussion rolla granular stick techniquecan be applied to any surface, pitched or unpitched.
In the literature of acoustics and signal processing, many terms refer to
similar microsonic phenomena: acoustic quantum, sonal atom, grain, glisson,
grainlet, trainlet, Gaussian elementary signal, Gaussian pulse, short-time segment,
sliding window, microarc, voicel, Coiet, symmlet, Gabor atom, Gabor wavelet,
gaborette, wavelet, chirplet, Lienard atom, FOF, FOG, wave packet, Vosim pulse,
time-frequency atom, pulsar, waveset, impulse, toneburst, tone pip, acoustic pixel,
and window function pulse are just a few. These phenomena, viewed in their
mathematical dual spacethe frequency domaintake on a dierent set of
names: kernel, logon, and frame, for example.
Perception of Microsound
Microevents last only a very short time, near to the threshold of auditory perception. Much scientic study has gone into the perception of microevents.
Human hearing mechanisms, however, intertwine with brain functions, cognition, and emotion, and are not completely understood. Certain facts are clear.

22

Chapter 1

One cannot speak of a single time frame, or a time constant for the auditory
system (Gordon 1996). Our hearing mechanisms involve many dierent agents,
each of which operates on its own time scale (see gure 1.1). The brain integrates signals sent by various hearing agents into a coherent auditory picture.
Ear-brain mechanisms process high and low frequencies dierently. Keeping
high frequencies constant, while inducing phase shifts in lower frequencies,
causes listeners to hear a dierent timbre.
Determining the temporal limits of perception has long engaged psychoacousticians (Doughty and Garner 1947; Buser and Imbert 1992; Meyer-Eppler
1959; Winckel 1967; Whiteld 1978). The pioneer of sound quanta, Dennis
Gabor, suggested that at least two mechanisms are at work in microevent detection: one that isolates events, and another that ascertains their pitch. Human
beings need time to process audio signals. Our hearing mechanisms impose
minimum time thresholds in order to establish a rm sense of the identity and
properties of a microevent.
In their important book Audition (1992), Buser and Imbert summarize a large
number of experiments with transitory audio phenomena. The general result
from these experiments is that below 200 ms, many aspects of auditory perception change character and dierent modes of hearing come into play. The
next sections discuss microtemporal perception.
Microtemporal Intensity Perception
In the zone of low amplitude, short sounds must be greater in intensity than
longer sounds to be perceptible. This increase is about 20 dB for tone pips
of 1 ms over those of 100 ms duration. (A tone pip is a sinusoidal burst with
a quasi-rectangular envelope.) In general, subjective loudness diminishes with
shrinking durations below 200 ms.
Microtemporal Fusion and Fission
In dense portions of the Milky Way, stellar images appear to overlap, giving the eect of a
near-continuous sheet of light . . . The eect is a grand illusion. In reality . . . the nightime
sky is remarkably empty. Of the volume of space only 1 part in 10 21 [one part in a quintillion] is lled with stars. (Kaler 1997)

Circuitry can measure time and recognize pulse patterns at tempi in the range
of a gigahertz. Human hearing is more limited. If one impulse follows less than
200 ms after another, the onset of the rst impulse will tend to mask the second,

23

Time Scales of Music

a time-lag phenomenon known as forward masking, which contributes to the

illusion that we call a continuous tone.
The sensation of tone happens when human perception reaches attentional
limits where microevents occur too quickly in succession to be heard as discrete
events. The auditory system, which is nonlinear, reorganizes these events into
a group. For example, a series of impulsions at about 20 Hz fuse into a continuous tone. When a fast sequence of pitched tones merges into a continuous
``ripple,'' the auditory system is unable to successfully track its rhythm. Instead,
it simplies the situation by interpreting the sound as a continuous texture. The
opposite eect, tone ssion, occurs when the fundamental frequency of a tone
descends into the infrasonic frequencies.
The theory of auditory streams (McAdams and Bregman 1979) aims to explain the perception of melodic lines. An example of a streaming law is: the
faster a melodic sequence plays, the smaller the pitch interval needed to split it
into two separately perceived ``streams.'' One can observe a family of streaming
eects between two alternating tones A and B. These eects range from coherence (the tones A and B form a single percept), to roll (A dominates B), to
masking (B is no longer perceived).
The theory of auditory streaming was an attempt to create a psychoacoustic
basis for contrapuntal music. A fundamental assumption of this research was
that ``several musical dimensions, such as timbre, attack and decay transients,
and tempo are often not specied exactly by the composer and are controlled
by the performer'' (McAdams and Bregman 1979). In the domain of electronic
music, such assumptions may not be valid.
Microtemporal Silence Perception
The ear is quite sensitive to intermittencies within pure sine waves, especially in
the middle range of frequencies. A 20 ms uctuation in a 600 Hz sine wave,
consisting of a 6.5 ms fade out, a 7 ms silent interval, and a 6.5 ms fade in,
breaks the tone in two, like a double articulation. A 4 ms interruption, consisting of a 1 ms fade out, a 2 ms silent interval, and a 1 ms fade in, sounds like
a transient pop has been superimposed on the sine wave.
Intermittencies are not as noticeable in complex tones. A 4 ms interruption is
not perceptible in pink noise, although a 20 ms interruption is.
In intermediate tones, between a sine and noise, microtemporal gaps less
than 10 ms sound like momentary uctuations in amplitude or less noticeable
transient pops.

24

Chapter 1

Microtemporal Pitch Perception

Studies by Meyer-Eppler show that pitch recognition time is dependent on frequency, with the greatest pitch sensitivity in the mid-frequency range between
1000 and 2000 Hz, as the following table (cited in Butler 1992) indicates.
Frequency in Hz
Minimum duration in ms

100
45

500
26

1000
14

5000
18

Doughty and Garner (1947) divided the mechanism of pitch perception into
two regions. Above about 1 kHz, they estimated, a tone must last at least 10 ms
to be heard as pitched. Below 1 kHz, at least two to three cycles of the tone are
needed.
Microtemporal Auditory Acuity
We feel impelled to ascribe a temporal arrangement to our experiences. If b is later than a
and g is later than b, then g is also later than a. At rst sight it appears obvious to assume
that a temporal arrangement of events exists which agrees with the temporal arrangement
of experiences. This was done unconsciously until skeptical doubts made themselves felt.
For example, the order of experiences in time obtained by acoustical means can dier from
the temporal order gained visually . . . (Einstein 1952)

Green (1971) suggested that temporal auditory acuity (the ability of the ear to
detect discrete events and to discern their order) extends down to durations as
short as 1 ms. Listeners hear microevents that are less than about 2 ms in duration as a click, but we can still change the waveform and frequency of these
events to vary the timbre of the click. Even shorter events (in the range of
microseconds) can be distinguished on the basis of amplitude, timbre, and spatial position.
Microtemporal Preattentive Perception
When a person glimpses the face of a famous actor, snis a favorite food, or hears the voice
of a friend, recognition is instant. Within a fraction of a second after the eyes, nose, ears,
tongue or skin is stimulated, one knows the object is familiar and whether it is desirable or
dangerous. How does such recognition, which psychologists call preattentive perception,
happen so accurately and quickly, even when the stimuli are complex and the context in
which they arise varies? (Freeman 1991)

One of the most important measurements in engineering is the response of a

system to a unit impulse. It should not be surprising to learn that auditory

25

Time Scales of Music

neuroscientists have sought a similar type of measurement for the auditory

system. The impulse response equivalents in the auditory system are the auditory evoked potentials, which follow stimulation by tone pips and clicks.
The rst response in the auditory nerve occurs about 1.5 ms after the initial
stimulus of a click, which falls within the realm of preattentive perception
(Freeman 1995). The mechanisms of preattentive perception perform a rapid
analysis by an array of neurons, combining this with past experience into a
wave packet in its physical form, or a percept in its behavioral form. The neural
activities sustaining preattentive perception take place in the cerebral cortex.
Sensory stimuli are preanalyzed in both the pulse and wave modes in intermediate stations of the brain. As Freeman noted, in the visual system complex
operations such as adaptation, range compression, contrast enhancement,
and motion detection take place in the retina and lower brain. Sensory stimuli
activate feature extractor neurons that recognize specic characteristics.
Comparable operations have been described for the auditory cortex: the nal
responses to a click occur some 300 ms later, in the medial geniculate body of
the thalamus in the brain (Buser and Imbert 1992).
Microtemporal Subliminal Perception
Finally, we should mention subliminal perception, or perception without awareness. Psychological studies have tested the inuence of brief auditory stimuli
on various cognitive tasks. In most studies these take the form of verbal hints to
some task asked of the listener. Some evidence of inuence has been shown, but
the results are not clear-cut. Part of the problem is theoretical: how does subliminal perception work? According to a cognitive theory of Reder and Gordon
(1997), for a concept to be in conscious awareness, its activation must be above
a certain threshold. Magnitude of activation is partly a function of the exposure
duration of the stimulus. A subliminal microevent raises the activation of the
corresponding element, but not enough to reach the threshold. The brain's
``production rules'' cannot re without the elements passing threshold, but a
subliminal microevent can raise the current activation level of an element
enough to make it easier to re a production rule later.
The musical implications are, potentially, signicant. If the subliminal hints
are not fragments of words but rather musical cues (to pitch, timbre, spatial
position, or intensity) then we can embed such events at pivotal instants, knowing that they will contribute to a percept without the listener necessarily being
aware of their presence. Indeed this is one of the most interesting dimensions of
microsound, the way that subliminal or barely perceptible variations in the

26

Chapter 1

properties of a collection of microeventstheir onset time, duration, frequency,

waveform, envelope, spatial position, and amplitudelead to dierent aesthetic
perceptions.
Viewing and Manipulating the Microtime Level
Microevents touch the extreme time limits of human perception and performance. In order to examine and manipulate these events uidly, we need digital
audio ``microscopes''software and hardware that can magnify the micro time
scale so that we can operate on it.
For the serious researcher, the most precise strategy for accessing the micro
time scale is through computer programming. Beginning in 1974, my research
was made possible by access to computers equipped with compiler software
and audio converters. Until recently, writing one's own programs was the only
possible approach to microsound synthesis and transformation.
Many musicians want to be able to manipulate this domain without the total
immersion experience that is the lifestyle of software engineering. Fortunately,
the importance of the micro time scale is beginning to be recognized. Any sound
editor with a zoom function that proceeds down to the sample level can view
and manipulate sound microstructure (gure 1.4).
Programs such as our Cloud Generator (Roads and Alexander 1995),
oer high-level controls in the micro time domain (see appendix A). Cloud
Generator's interface directly manipulates the process of particle emission,
controlling the ow of many particles in an evolving cloud. Our more recent
PulsarGenerator, described in chapter 4, is another example of a synthetic
particle generator.
The perceived result of particle synthesis emerges out of the interaction of
parameter evolutions on a micro scale. It takes a certain amount of training to
learn how operations in the micro domain translate to acoustic perceptions on
higher levels. The grain duration parameter in granular synthesis, for example,
has a strong eect on the perceived spectrum of the texture.
This situation is no dierent from other well-known synthesis techniques.
Frequency modulation synthesis, for example, is controlled by parameters such
as carrier-to-modulator ratios and modulation indexes, neither of which are
direct terms of the desired spectrum. Similarly, physical modeling synthesis is
controlled by manipulating the parameters that describe the parts of a virtual instrument (size, shape, material, coupling, applied force, etc.), and not the sound.
One can imagine a musical interface in which a musician species the desired
sonic result in a musically descriptive language which would then be translated

27

Time Scales of Music

Figure 1.4 Viewing the micro time scale via zooming. The top picture is the waveform
of a sonic gesture constructed from sound particles. It lasts 13.05 seconds. The middle
image is a result of zooming in to a part of the top waveform (indicated by the dotted
lines) lasting 1.5 seconds. The bottom image is a microtemporal portrait of a 10 millisecond fragment at the beginning of the top waveform (indicated by the dotted lines).

into particle parameters and rendered into sound. An alternative would be to

specify an example: ``Make me a sound like this (soundle), but with less
vibrato.'' This is a challenging task of parameter estimation, since the system
would have to interpret how to approximate a desired result. For more on the
problems of parameter estimation in synthesis see Roads (1996).
Do the Particles Really Exist?
In the 1940s, the physicist Dennis Gabor made the assertion that all sound
even continuous tonescan be considered as a succession of elementary particles of acoustic energy. (Chapter 2 summarizes this theory.) The question then
arises: do sound particles really exist, or are they merely a theoretical con-

28

Chapter 1

struction? In certain sounds, such as the taps of a slow drum roll, the individual
particles are directly perceivable. In other sounds, we can prove the existence of
a granular layer through logical argument.
Consider the whole number 5. This quantity may be seen as a sum of subquantities, for example 1 1 1 1 1, or 2 3, or 4 1, and so on. If we
take away one of the subquantities, the sum no longer is 5. Similarly, a continuous tone may be considered as a sum of subquantitiesas a sequence of overlapping grains. The grains may be of arbitrary sizes. If we remove any grain,
the signal is no longer the same. So clearly the grains exist, and we need all of
them in order to constitute a complex signal. This argument can be extended
to explain the decomposition of a sound into any one of an innite collection
of orthogonal functions, such as wavelets with dierent basis functions, Walsh
functions, Gabor grains, and so on.
This logic, though, becomes tenuous if it is used to posit the preexistence (in
an ideal Platonic realm) of all possible decompositions within a whole. For example, do the slices of a cake preexist, waiting to be articulated? The philosophy of mathematics is littered with such questions (Castonguay 1972, 1973).
Fortunately it is not our task here to try to assay their signicance.
Heterogeneity in Sound Particles
The concept of heterogeneity or diversity of sound materials, which we have
already discussed in the context of the sound object time scale, also applies to
other time scales. Many techniques that we use to generate sound particles assign to each particle a unique identity, a precise frequency, waveform, duration,
amplitude morphology, and spatial position, which then distinguishes it from
every other particle. Just as certain sound objects may function as singularities,
so may certain sound particles.

Sampled Time Scale

Below the level of microtime stands the sampled time scale (gure 1.5). The
electronic clock that drives the sampling process establishes a time grid. The
spacing of this grid determines the temporal precision of the digital audio
medium. The samples follow one another at a xed time interval of 1= fS , where
fS is the sampling frequency. When fS 44:1 kHz (the compact disc rate),
the samples follow one another every 22.675 millionths of a second (msec).

29

Time Scales of Music

Figure 1.5 Sample points in a digital waveform. Here are 191 points spanning a 4.22 ms
time interval. The sampling rate is 44.1 kHz.

The atom of the sample time scale is the unit impulse, the discrete-time counterpart of the continuous-time Dirac delta function. All samples should be considered as time-and-amplitude-transposed (delayed and scaled) instances of
the unit impulse.
The interval of one sample period borders near the edge of human audio
perception. With a good audio system one can detect the presence of an individual high-amplitude sample inserted into a silent stream of zero-valued samples. Like a single pixel on a computer screen, an individual sample oers little.
Its amplitude and spatial position can be discerned, but it transmits no sense of
timbre and pitch. Only when chained into sequences of hundreds do samples
oat up to the threshold of timbral signicance. And still longer sequences of
thousands of samples are required to represent pitched tones.
Sound Composition with Individual Sample Points
Users of digital audio systems rarely attempt to deal with individual sample
points, which, indeed, only a few programs for sound composition manipulate
directly. Two of these are G. M. Koenig's Sound Synthesis Program (SSP) and

30

Chapter 1

Herbert Brun's Sawdust program, both developed in the late 1970s. Koenig and
Brun emerged from the Cologne school of serial composition, in which the interplay between macro- and microtime was a central aesthetic theme (Stockhausen 1957; Koenig 1959; Maconie 1989). Brun wrote:
For some time now it has become possible to use a combination of analog and digital
computers and converters for the analysis and synthesis of sound. As such a system will
store or transmit information at the rate of 40,000 samples per second, even the most
complex waveforms in the audio-frequency range can be scanned and registered or be
recorded on audio tape. This . . . allows, at last, the composition of timbre, instead of with
timbre. In a sense, one may call it a continuation of much which has been done in the electronic music studio, only on a dierent scale. The composer has the possibility of extending
his compositional control down to elements of sound lasting only 1/20,000 of a second.
(Brun 1970)

Koenig's and Brun's synthesis programs were conceptually similar. Both

represented a pure and radical approach to sound composition. Users of these
programs stipulated sets of individual time and amplitude points, where each
set was in a separate le. They then specied logical operations such as linking,
mingling, and merging, to map from a time-point set to an amplitude-point set
in order to construct a skeleton of a waveform fragment. Since these points
were relatively sparse compared to the number of samples needed to make a
continuous sound, the software performed a linear interpolation to connect intermediate amplitude values between the stipulated points. This interpolation,
as it were, eshed out the skeleton. The composer could then manipulate the
waveform fragments using logical set theory operations to construct larger and
larger waveforms, in a process of hierarchical construction.
Koenig was explicit about his desire to escape from the traditional computergenerated sounds:
My intention was to go away from the classical instrumental denitions of sound in terms
of loudness, pitch, and duration and so on, because then you could refer to musical elements
which are not necessarily the elements of the language of today. To explore a new eld of
sound possibilities I thought it best to close the classical descriptions of sound and open up
an experimental eld in which you would really have to start again. (Roads 1978b)

Iannis Xenakis proposed a related approach (Xenakis 1992; Homann 1994,

1996, 1997). This involves the application of sieve theory to the amplitude and
time dimensions of a sound synthesis process. As in his Gendyn program, the
idea is to construct waveforms from fragments. Each fragment is bounded by
two breakpoints. Between the breakpoints, the rest of the waveform is lled in

31

Time Scales of Music

by interpolation. Whereas in Gendyn the breakpoints are calculated from a

nonlinear stochastic algorithm, in sieve theory the breakpoints would be calculated according to a partitioning algorithm based on sieved amplitude and time
dimensions.
Assessment of Sound Composition with Samples
To compose music by means of logical operations on samples is a daunting
task. Individual samples are subsymbolicperceptually indistinguishable from
one another. It is intrinsically dicult to string together samples into meaningful music symbols. Operations borrowed from set theory and formal logic do
not take into account the samples' acoustical signicance. As Koenig's statement above makes clear, to compose intentionally a graceful melodic gure, a
smooth transition, a cloud of particles, or a polyphonic texture requires extraordinary eort, due to the absence of acoustically relevant parameters for building higher-level sound structures. Users of sample-based synthesis programs
must be willing to submit to the synthesis algorithm, to abandon local control,
and be satised with the knowledge that the sound was composed according
to a logical process. Only a few composers took up interest in this approach,
and there has not been a great deal of experimentation along these lines since
the 1970s.

Subsample Time Scale

A digital audio system represents waveforms as a stream of individual samples
that follow one another at a xed time interval (1= fS , where fS is the sampling
frequency). The subsample time scale supports uctuations that occur in less
than two sampling periods. Hence this time scale spans a range of minuscule
durations measured in nanoseconds and extending down to the realm of innitesimal intervals.
To stipulate a sampling frequency is to x a strict threshold between a subsample and the sample time scale. Frequencies above this thresholdthe
Nyquist frequency (by denition: fS =2)cannot be represented properly by a
digital audio system. For the standard compact disc sampling rate of 44.1 kHz,
the Nyquist frequency is 22.05 kHz. This means that any wave uctuation
shorter than two samples, or 45 msec, is relegated to the subsample domain. The
96 kHz sampling rate standard reduces this interval to 20.8 msec.

32

Chapter 1

The subsample time scale encompasses an enormous range of phenomena.

Here we present ve classes of subsample phenomena, from the real and perceptible to the ideal and imperceptible: aliased artefacts, ultrasounds, atomic
sounds, and the Planck interval.
Aliased Artefacts
In comparison with the class of all time intervals, the class of perceptible
audio periods spans relatively large time intervals. In a digital audio system, the
sample period is a threshold separating all signal uctuations into two classes:
those whose frequencies are low enough to be accurately recorded and those
whose frequencies are too high to be accurately recorded. Because a frequency
is too high to be recorded does not mean that it is invisible to the digital recorder. On the contrary, subsample uctuations, according to the theorem of
Nyquist (1928), record as aliased artefacts. Specically, if the input frequency is
higher than half the sampling frequency, then:
aliased frequency sampling frequency input frequency
Thus if the sampling rate is 44.1 kHz, an input frequency of 30 kHz is
reected down to the audible 11.1 kHz. Digital recorders must, therefore, attempt to lter out all subsample uctuations in order to eliminate the distortion
caused by aliased artefacts.
The design of antialiasing lters has improved in the past decade. Current
compact disc recordings are eectively immune from aliasing distortion. But
the removal of all information above 22.05 kHz poses problems. Many people
hear detail (referred to as air) in the region above 20 kHz (Koenig 1899; Neve
1992). Rigorous scientic experiments have conrmed the eects, from both
physiological and subjective viewpoints, of sounds above 22 kHz (Oohashi et
al. 1991; Oohashi et al. 1993). Furthermore, partials in the ultrasonic region
interact, resulting in audible subharmonics and air. When the antialiasing lter
removes these ultrasonic interactions, the recording loses detail.
Aliasing remains a pernicious problem in sound synthesis. The lack of frequency headroom in the compact disc standard rate of 44.1 kHz opens the door
to aliasing from within the synthesis algorithm. Even common waveforms cause
aliasing when extended beyond a narrow frequency range. Consider these cases
of aliasing in synthesis:
1. A band-limited square wave made from sixteen odd-harmonic components
causes aliasing at fundamental frequencies greater than 760 Hz.

33

Time Scales of Music

2. An additive synthesis instrument with thirty-two harmonic partials generates

aliased components if the fundamental is higher than 689 Hz (approximately
E5).
3. The partials of a sampled piano tone A-sharp2 (116 Hz) alias when the tone
is transposed an octave and a fth to F4 (349 Hz).
4. A sinusoidal frequency modulation instrument with a carrier-to-modulator
ratio of 1 : 2 and a fundamental frequency of 1000 Hz aliases if the modulation index exceeds 7. If either the carrier or modulator is a non-sinusoidal
waveform then the modulation index must typically remain less than 1.
As a consequence of these hard limits, synthesis instruments require preventative measures in order to eliminate aliasing distortion. Commercial instruments lter their waveforms and limit their fundamental frequency range. In
experimental software instruments, we must introduce tests and constrain the
choice of waveforms above certain frequencies.
The compact disc sampling rate of 44.1 kHz rate is too low for high-delity
music synthesis applications. Fortunately, converters operating at 96 kHz are
becoming popular, and sampling rates up to 192 kHz also are available.
Ultrasonic Loudspeakers
Even inaudible energy in the ultrasonic frequency range can be harnessed for
audio use. New loudspeakers have been developed on the basis of acoustical
heterodyning (American Technology Corporation 1998; Pompei 1998). This
principle is based on a phenomenon observed by Helmholtz. When two sound
sources are positioned relatively closely together and are of a suciently high
amplitude, two new tones appear: one lower and one higher than either of the
original tones. The two new combination tones correspond to the sum and the
dierence of the two original tones. For example, if one were to emit 90 kHz
and 91 kHz into the air, with sucient energy, one would produce the sum
(181 kHz) and the dierence (1 kHz), the latter being in the range of human
hearing. Reporting that he could also hear summation tones (whose frequency
is the sum, rather than the dierence, of the two fundamental tones), Helmholtz argued that the phenomenon had to result from a nonlinearity of air molecules. Air molecules begin to behave nonlinearly (to heterodyne) as amplitude
increases. Thus, a form of acoustical heterodyning is realized by creating difference frequencies from higher frequency waves. In air, the eect works in

34

Chapter 1

such a way that if an ultrasonic carrier is increased in amplitude, a dierence

frequency is created. Concurrently, the unused sum frequency diminishes in
loudness as the carrier's frequency increases. In other words, the major portion
of the ultrasonic energy transfers to the audible dierence frequency.
Unlike regular loudspeakers, acoustical heterodyning loudspeakers project
energy in a collimated sound beam, analogous to the beam of light from a
ashlight. One can direct an ultrasonic emitter toward a wall and the listener
will perceive the sound as coming from a spot on that wall. For a direct sound
beam, a listener standing anywhere in an acoustical environment is able to
point to the loudspeaker as the source.
Atomic Sound: Phonons and Polarons
As early as 1907, Albert Einstein predicted that ultrasonic vibration could
occur on the scale of atomic structure (Cochran 1973). The atoms in crystals,
he theorized, take the form of a regular lattice. A one-dimensional lattice
resembles the physical model of a taut stringa collection of masses linked
by springs. Such a model may be generalized to other structures, for example
three-dimensional lattices. Lattices can be induced to vibrate ultrasonically,
subjected to the proper force, turning them into high-frequency oscillators. This
energy is not continuous, however, but is quantized by atomic structure into
units that Einstein called phonons, by analogy to photonsthe quantum units
of light. It was not until 1913 that regular lattices were veried experimentally
as being the atomic structure of crystals. Scientists determined that the frequency of vibration depends on the mass of the atoms and the nature of the
interatomic forces. Thus the lower the atomic weight, the higher the frequency
of the oscillator (Stevenson and Moore 1967). Ultrasonic devices can generate
frequencies in the trillions of cycles per second.
Complex sound phenomena occur when phononic energy collides with other
phonons or other atomic particles. When the sources of excitation are multiple
or the atomic structure irregular, phonons propagate in cloud-like swarms
called polarons (Pines 1963). Optical energy sources can induce or interfere with
mechanical vibrations. Thus optical photons can scatter acoustic phonons. For
example, laser-induced lattice vibrations can change the index of refraction in a
crystal, which changes its electromagnetic properties. On a microscopic scale,
optical, mechanical, and electromagnetic quanta are interlinked as elementary
excitations.

35

Time Scales of Music

Laser-induced phonic sound focuses the beams from two lasers with a small
wavelength dierence onto a crystal surface. The dierence in wavelength
causes interference, or beating. The crystal surface shrinks and expands as
this oscillation of intensity causes periodic heating. This generates a wave that
propagates through the medium. The frequency of this sound is typically in the
gigahertz range, with a wavelength of the order of 1 micron. Because of the
small dimensions of the heated spot on the surface, the wave in the crystal has
the shape of a directional beam. These sound beams can be used as probes, for
example, to determine the internal features of semiconductor crystals, and to
detect faults in their structure.
One of the most important properties of laser-induced phononic sound is that
it can be made coherent (the wave trains are phase-aligned), as well as monochromatic and directional. This makes possible such applications as acoustic
holography (the visualization of acoustic phenomena by laser light). Today the
study of phononic vibrations is an active eld, nding applications in surface
acoustic wave (SAW) lters, waveguides, and condensed matter physics.
At the Physical Limits: The Planck Time Interval
Sound objects can be subdivided into grains, and grains into samples. How far
can this subdivision of time continue? Hawking and Penrose (1996) have suggested that time in the physical universe is not innitely divisible. Specically,
that no signal uctuation can be faster than the quantum changes of state in
subatomic particles, which occur at close to the Planck scale. The Planck scale
stands at the extreme limit of the known physical world, where current concepts
of space, time, and matter break down, where the four forces unify. It is the
exceedingly small distance, related to an innitesimal time span and extremely
high energy, that emerges when the fundamental constants for gravitational
attraction, the velocity of light, and quantum mechanics join (Hawking and
Penrose 1996).
How much time does it take light to cross the Planck scale? Light takes about
3.3 nanoseconds (3:3  1010 ) to traverse 1 meter. The Planck time interval is
the time it takes light to traverse the Planck scale. Up until recently, the Planck
scale was thought to be 1033 meter. An important new theory puts the gure at
a much larger 1019 meter (Arkani-Hamed et al. 2000). Here, the Planck time
interval is 3:3  1028 seconds, a tiny time interval. One could call the Plank
time interval a kind of ``sampling rate of the universe,'' since no signal uctuation can occur in less than the Planck interval.

36

Chapter 1

If the ow of time stutters in discrete quanta corresponding to fundamental

physical constants, this poses an interesting conundrum, recognized by Iannis
Xenakis:
Isn't time simply an epiphenomenal notion of a deeper reality? . . . The equations of
Lorentz-Fitzgerald and Einstein link space and time because of the limited velocity of light.
From this it follows that time is not absolute . . . It ``takes time'' to go from one point to
another, even if that time depends on moving frames of reference relative to the observer.
There is no instantaneous jump from one point to another in space, much less spatial
ubiquitythat is, the simultaneous presence of an event or object everywhere in space. To
the contrary, one posits the notion of displacement. Within a local reference frame, what
does displacement signify? If the notion of displacement were more fundamental than that
of time, one could reduce all macro and micro cosmic transformations to weak chains of
displacement. Consequently . . . if we were to adhere to quantum mechanics and its implications, we would perhaps be forced to admit the notion of quantied space and its corollary, quantied time. But what could a quantied time and space signify, a time and space
in which contiguity would be abolished. What would the pavement of the universe be if
there were gaps between the paving stones, inaccessible and lled with nothing? (Xenakis
1989)

Innitesimal Time Scale

Besides the innite-duration sinusoids of Fourier theory, mathematics has created other ideal, innite-precision boundary quantities. One class of ideal phenomena that appears in the theory of signal processing is the mathematical
impulse or delta (q) function. Delta functions represent innitely brief intervals
of time. The most important is the Dirac delta function, formulated for the
theory of quantum mechanics. Imagine the time signal shown in gure 1.6a, a
narrow pulse of height 1=b and width b, centered on t 0. This pulse, xt, is
zero at all times jtj > b=2. For any nonzero value of b, the integral of xt is
unity. Imagine that b shrinks to a duration of 0. Physically this means that
the pulse's height grows and the interval of integration (the pulse's duration)
becomes very narrow. The limit of xt as b ! 0 is shown in gure 1.6b. This
shows that the pulse becomes an innitely high spike of zero width, indicated as
qt, the Dirac delta function. The two signicant properties of the q function
are: (1) it is zero everywhere except at one point, and (2) it is innite in amplitude at this point, but approaches innity in such a way that its integral is
unitya curious object!

37

Time Scales of Music

Figure 1.6 Comparison of a pulse and the Dirac delta function. (a) A narrow pulse of
height 1=b and width b, centered on t 0. (b) The Dirac delta function.

38

Chapter 1

The main application of the q function in signal processing is to bolster the

mathematical explanation of the process of sampling. When a q function occurs
inside an integral, the value of the integral is determined by nding the location
of the impulse and then evaluating the integrand at that location. Since the q is
innitely brief, this is equivalent to sampling the function being integrated.
Another interesting property of the q function is that its Fourier transform,
jej2pft j 1
for any real value of t. In other words, the spectrum of an innitely brief impulse is innite (Nahin 1996).
We see here a profound law of signal processing, which we will encounter
repeatedly in this thesis, that duration and spectrum are complementary quantities. In particular, the shorter a signal is, the broader is its spectrum. Later we
will see that one can characterize various signal transformations by how they
respond to the q function and its discrete counterpart, the unit impulse.
The older Kronecker delta is an integer-valued ideal impulse function. It is
dened by the properties

0 m0n
qm; n
1 mn
The delta functions are dened over a continuous and innite domain. The
section on aliased artefacts examines similar functions in the discrete sampled
domain.

Outside Time Music

Musical structure can exist, in a sense, ``outside'' of time (Xenakis 1971, 1992).
By this, we mean abstract structuring principles whose denition does not imply
a temporal order. A scale, for example, is independent of how a composer uses
it in time. Myriad precompositional strategies, and databases of material could
also be said to be outside time.
A further example of an outside time structure is a musical instrument.
The layout of keys on a piano gives no hint of the order in which they will
be played. Aleatoric compositions of the 1950s and 1960s, which left various
parameters, including the sequence of events to chance, were also outside time
structures.

39

Time Scales of Music

Today we see installations and virtual environments in which sounds occur in

an order that depends on the path of the person interacting with the system. In
all of these cases, selecting and ordering the material places it in time.

The Size of Sounds

Sounds form in the physical medium of aira gaseous form of matter. Thus,
sound waves need space to form. Just as sounds exist on dierent time scales,
so they take shape on dierent scales of space. Every sound has a threedimensional shape and size, which is its diusion or dispersion pattern over time.
Since the wavelength of a high frequency sound is short, high frequencies form
in small spaces. A low frequency waveform needs several meters to unfold. The
temporal and the spatial morphologies of a sound intertwine. A sound's duration, frequency, amplitude, and pattern of radiation from its source all contribute to its physical form, as does the space in which the sound manifests.
The duration of a sound is an important determinant of physical shape, especially in the open air. A long-duration sound is long in spatial extent, spanning the entire distance from the source to the point at which its energy is
completely absorbed. Short-duration sounds, on the contrary, are thin in spatial
extent, disappearing from their point of origin quickly. The wave of a shortduration sound occupies a thin band of air, although the uctuations that it
carries may travel great distances if it is loud enough.
Today we have accurate measurements of the speed of sound waves in a variety of media (Pierce 1994). The accepted value for the speed of sound in dry
air is 331.5 meters/second. Thus a 20 Hz acoustical wave requires no less than
16.5 meters (54.13 feet) to unfold without obstruction. Obstructions such as
walls cause the wave to reect back on itself, creating phase cancellation eects.
A high-frequency waveform at 20 kHz has a period of only 1/20,000th of a
second. This takes only 1.65 cm to form. The ear is very sensitive to the time of
arrival of sounds from dierent spatial positions. Thus, even a minor dierence
in the distance of the listener from two separate sources will skew the spatial
images.
The most important determinant of a sound's size is its amplitude. Very loud
sounds (such as atmospheric thunder and other explosions) travel far. As they
travel, the air gradually absorbs the high frequencies, so that only the low frequencies reach great distances.

40

Chapter 1

Summary
Particle physics seeks to nd a simple and orderly pattern to the behavior of matter on the
atomic and subatomic level. To this end, large particle accelerators are built, acting like
giant microscopes that zoom down through the atom . . . Astronomers build equally complex devicestelescopes and observatories. These gather data from distant clusters of
galaxies, all the way out to the rim of the cosmos . . . We are seeing here a convergence
between particle physics and cosmology. The instruments, and even the stated objectives,
are dierent, but the languages draw closer. The laws of nature that control and order the
microscopic world, and those that determined the creation and evolution of the universe, . . .
are beginning to look identical. (Lederman and Schramm 1995)

Projecting time horizontally, and amplitude vertically, the concept of nil duration corresponds to a zero-dimensional point on the time-amplitude plane. This
point zero is mute: no ux of energy can occur in the absence of a time window.
In that ideal world experienced only by the gods of mathematics, the delta
function qt breaks the monotony with an instantaneous impulse that is born
and dies within the most innitesimal window beyond point zero.
Our mundane digital domain is a discrete approximation to the ideal realm
of innitesimal time. In the digital domain, the smallest event has a duration
equivalent to the period of the sampling frequency. This sound atom, the sample period, is the grid that quantizes all time values in an audio signal. Any
curve inscribed on the amplitude-versus-time plane must synchronize to this
grid. Individual samples remain subsymbolic. Like the woven threads of canvas
holding paint in place, their presence is a necessity, even if we can see them only
in the aggregate.
As the window of time expands, there is a possibility for chaotic uctuation,
periodic repetition, echoes, tone, noise, and measured silence. Each additional
instant of time accrues new possibilities.
Microsonic particles can be likened to molecules built from atomic samples.
To view this level of detail, we rely on the tools of sound analysis and display.
Under this scrutiny, remarkable patterns emerge and we gain new insight into
sound structure. These images show the hidden morphologies of elementary
sound molecules (gure 1.7).
Molecular materials alter the terrain of composition. Pliant globules can be
molded into arbitrary object morphologies. The presence of mutating sound
objects suggests a uid approach to compositional mesostructure, spawning
rivulets, streams, and clouds as well as discrete events. The package for all these

41

Time Scales of Music

Figure 1.7 Image of a grain in the time-domain (top) and its frequency-domain counterpart (bottom).

musical structures, the macroform, can be tailored with high exibility and
precision in a sound mixing program.
It is necessary to see music over a broad range of time scales, from the innitesimal to the supra scale (Christensen 1996). Not all musicians are prepared
to view musical time from such a comprehensive perspective, however, and
it may well take decades for this perspective to lter into our general musical
vocabulary.

This page intentionally left blank

The History of Microsound from Antiquity

to the Analog Era

Waves versus Particles: Early Concepts of Microsound

Optical Wave versus Particle Debate
Acoustical Wave versus Particle Debate
Waves versus Particles: a Contemporary Perspective
The Modern Concept of Microsound
Temporal Continuity in Perception
The Gabor Matrix
Electro-optical and Electromechanical Sound Granulation
Meyer-Eppler
Moles
Wiener
Theory and Experiments of Xenakis
Organization of Analogique B
Problems with a Constant Microtime Grid
Analog Impulse Generators
Stockhausen's Temporal Theory
How Time Passes
The Unity of Musical Time
Assessment of Stockhausen's Temporal Theory
Other Assessments of Stockhausen's Temporal Theory
Microsound in the Analog Domain

44

Chapter 2

Microsound in the Instrumental Domain

Summary
Musical ideas are prisoners, more than one might believe, of musical devices.
Schaeer (1977, pp. 1617)

The evolution of sound synthesis has always been interwoven with the engines of acoustic emission, be they mechanoacoustic, electromechanical, electrooptical, analog electronic, or digital. The current state of music technology has
been arrived at through decades of laboratory experimentation. If we are to
benet from this legacy, we must revisit the past and recover as much knowledge as we can.
Table 2.1 lists electric and electronic music instruments developed in the period 18991950. The rst column names each instrument. The second column
shows the date of their rst public demonstration (rather than the date of
their conception). Before 1950, almost all instruments were designed for live
performance. After 1950, the technology of recording changed the nature of
electronic music, ushering in the era of the tape-based electronic music studio.
Electronic instruments invented before 1950 represented a wave-oriented approach to synthesis, as opposed to a particle-oriented approach. Gabor's experiments in the late 1940s signaled the beginning of a new era in synthesis.
This chapter explores the ancient philosophical debate between waves and
particles. It then presents the modern history of microsound synthesis, continuing through the era of analog electronics. Chapter 7 continues this story by
recounting the history of early experiments in microsound synthesis by digital
computer.

Waves versus Particles: Early Concepts of Microsound

To view the microacoustical domain is to confront a scientic dilemma that
has confounded physicists for centuries: the wave versus the particle nature of
signal energy. Debates concerning electromagnetic signals (such as light) have
motivated most scientic inquiries. But much of what has been discovered about
these signals applies to soundthe domain of mechanoacoustic vibrations as
well. We will briey look at the debate in both domains, optics and acoustics.

45

The History of Microsound

Table 2.1 Electric and electronic musical instruments: 18991950

Instrument

Date of
demonstration

Inventor

Notes
Early electric keyboard instrument
Electromagnetic instrument

Rotating tone wheels to generate current, the

current drove metallic resonating bars
Antenna instrument played with hands in air;
based on heterodyne tone generator
Heterodyne tone generator with lter
Sharp attack, inductance-controlled keyboard
instrument
Improved Electrophon with keyboard
1200 divisions per octave, designed for studies
in melody and harmony
Polyphonic, based on vacuum-tube oscillators

Singing Arc
Choralcello Electric
Organ

1899
1903

Telharmonium
Audio oscillator and
Audion Piano
Synthetic Tone Musical
Instrument
Thereminovox

1906
1915

W. Duddell
Farrington,
C. Donahue, and
A. Homan
T. Cahill
L. De Forest

1918

S. Cabot

1920

L. Theremin

Electrophon
Staccatone

1921
1923

J. Mager
H. Gernsback

Sphaerophon
Electronic Harmonium

1926
1926

Pianorad
Violen

1926
c. 1926

Light Siren
Illuminovox
SuperPiano
Electric guitar
prototype
Electronic Violin

c. 1926
1926
1927
1927

J. Mager
L. Theremin and
?. Rzhevkin
H. Gernsback
W. Gurov and
?. Volynken
Kovakenko
L. Theremin
E. Spielmann
Les Paul

1927

E. Zitzmann-Zirini

Spielman Electric Piano

Harp
Ondes Martenot
Dynaphon
Hellertion

1928

J. Bethenod

1928
1928
1929

Crea-tone

1930

M. Martenot
R. Bertrand
B. Helberger and
P. Lertes
S. Cooper

Givelet-Coupleaux
organ

1930

J. Givelet and
E. Coupleaux

Rotating tone generators, massive synthesizer

First vacuum-tube instrument

Rotating optical disks and photocell detectors

Electro-optical projector with rotating disc
``Light-chopper'' instrument
Solid body construction with electromagnetic
pickups
Space control of pitch like the Theremin, but
switched control of volume
Microphone and speaker feedback to sustain
oscillations
First of many versions
Multivibrator oscllator
Vacuum-tube oscillator with feedback,
continuous linear controllers
Electric piano with feedback circuits for
sustain
Automated additive synthesis, oscillators
controlled by paper tape

46

Chapter 2
Table 2.1 (continued)

Instrument

Date of
demonstration

Inventor

Notes

Trautonium

1930

F. Trautwein

Neon-tube sawtooth tone generators,

resonance lters to emphasize formants

Magnetoelectric organ
Westinghouse organ

1930
1930

R. H. Ranger
R. Hitchcock

Ondium Pechadre

1930

Hardy-Goldwaithe
organ
Neo-Bechstein piano

1930

A. Hardy and
S. Brown
W. Nernst

Radiopiano
Trillion-tone Organ

1931
1931

Radiotone

1931

Rangertone Organ
Emicon

1931
1932

Gnome
Miessner Electronic
Piano
Rhythmicon

1932
1932

Mellertion
Electronde

1933
1933

Cellulophone
Elektroakustische Orgel

1933
1934

La Croix Sonore
Ethonium

1934
1934

H. Cowell,
L. Theremin,
B. Miessner
?
L. or M.
Taubman
P. Toulon
O. Vierling and
Kock
N. Oboukhov
G. Blake

Keyboard Theremin

1934

L. Theremin

Loar Vivatone

1934

L. Loar

1931

1932

Hiller
A. Lesti and
F. Sammis
Boreau
R. Ranger
N. Langer and
Hahnagyi
I. Eremeef
B. F. Miessner

Research instrument based on vacuum tube

oscillators
Theremin-like instrument with a volume key
instead of antenna
Electro-optical tone generators
Physics Institute, Berlin, piano with electrical
pickups instead of soundboard
Amplied piano
Electro-optical tone generators
String-induced radio-receiver tone generator
with lter circuits
Rotating tone wheels
Gas-discharge tube oscillator, controlled by
keyboard
Rotating electromagnetic tone wheels
88 electrostatic pickups
Complex rhythm machine with keyboard
10-division octave
Battery-powered, space control of pitch like
the Theremin, with volume pedal
Electro-optical tone generators
12 vacuum-tube master oscillators, other
pitches derived by frequency division
Heterodyning oscillator
Emulation of the Theremin heterodyne
oscillator
Bank of tone generators controlled by
traditional organ keyboard
A modied acoustic/electric guitar

47

The History of Microsound

Table 2.1 (continued)

Instrument

Date of
demonstration

Polytone

1934

Syntronic Organ

1934

Everett Orgatron

1934

Partiturphon
Hammond electric
organ
Photona

1935
1935

Inventor

Notes
Electro-optical tone generators

1935

A. Lesti and
F. Sammis
I. Eremeef and
L. Stokowski
F. Hoschke and
B. Miessner
J. Mager
L. Hammond and
B. Miessner
I. Eremeef

Variophone

1935

Y. Sholpo

Electrone

1935

Foerster Electrochord
Sonotheque

1936
1936

Compton Organ
Company
O. Vierling
L. Lavalee

Kraft-durch-Freude
Grosstonorgel

1936

Welte Light-Tone organ

National Dobro
VioLectric Violin and
Supro Guitar
Electric Hawaiian
guitar
Singing Keyboard

1936
1936

O. Vierling and
sta of HeinrichHertz-Institut,
Berlin
E. Welte
J. Dopyera

1936

L. Fender

1936

F. Sammis

Warbo Formant organ

1937

Oscillion

1937

Krakauer Electone
Melodium
Robb Wave organ

1938
1938
c. 1938

H. Bode and
C. Warnke
W. Swann and
W. Danforth
B. F. Miessner
H. Bode
M. Robb

Electro-optical tone generators; one-hour of

continuous variation
Amplied vibrating brass reeds
Five-voice Sphaerophon with three keyboards
Rotating tone generators
12 electro-optical tone generators, developed
at WCAU radio, Philadelphia
Photo-electric instrument in which the
musician draws the sound on sprocketed lm
Based on design of L. Bourn; electrostatic
rotary generators
Electromechanical piano
Coded performance instrument using
photoelectric translation of engraved grooves
Played at 1936 Olympic games

Electro-optical tone generators

Commercial instruments with electromagnetic
pickups
Commercial instrument with electromagnetic
pickups
Played electro-optical recordings, precursor
of samplers
Four-voice polyphonic, envelope shaping, key
assignment, two lters
Gas-discharge tube oscillator
Early electric piano
Touch-sensitive solo keyboard
Rotating electromagnetic tone generators

48

Chapter 2
Table 2.1 (continued)

Instrument

Date of
demonstration

Inventor

Notes

Sonor

c. 1939

?. Ananyev

Kaleidaphon
Allen organ
Neo Bechstein piano

1939
1939
1939

Amplied piano

1939

J. Mager
Jerome Markowitz
O. Vierling and
W. Nernst
B. Miessner

Moscow, ribbon controller on a horizontal

ngerboard; violin-like sound
``Kaleidoscopic'' tone mixtures
Vacuum-tube oscillators
First commercial version of the electric piano

Novachord

1939

Parallel Bandpass
Vocoder
Dynatone

1939

Voder speech
synthesizer
Violena
Emiriton

1939

Ekvodin

1940

V-8

c. 1940

Solovox

1940

W. Gurov
A. Ivanov and A.
Rimsky-Korsakov
A. Volodin,
Russia
A. Volodin,
Russia
L. Hammond

Univox

c. 1940

Univox Company

Multimonika

1940

Hohner GmbH

Ondioline

1941

Georges Jenny

Melotone

c. 1944

Hanert Electrical
Orchestra
Joergensen Clavioline

1945

Compton Organ
Company
J. Hanert

Rhodes Pre-Piano

1947

1939

1940
1940

1947

Hammond
Company
H. Dudley, Bell
Laboratories
B. Miessner,
A. Amsley
H. Dudley

M. Constant
Martin
H. Rhodes

Variable tonal quality depending on the

position of the pickups
Several tube oscillators, divide-down
synthesis, formant lters
Analysis and cross-synthesis
Electric piano
Voice model played by a human operator

Neon-tube oscillators

Monophonic vacuum-tube oscillator with

divide-down circuitry
Vacuum-tube sawtooth generator with diode
waveform shaper circuit
Lower manual is wind-blown, upper manual
has sawtooth generator
Multistable vibrator and lters, keyboard
mounted on springs for vibrato
Electrostatic rotary generators
Programmable performance controlled by
punched paper cards
Monophonic, three-octave keyboard
Metal tines amplied by electrostatic pickups

49

The History of Microsound

Table 2.1 (continued)

Instrument

Date of
demonstration

Inventor

Notes

Wurlitzer
Company
Conn Organ
Company
Hugh LeCaine

Based on the Orgatron reed design, later

modied according to B. Miessner's patents
Individual oscillators for each key

Wurlitzer electronic
organ
Conn Organ

1947

Electronic Sackbut

1948

Free Music Machine

1948

Mixturtrautonium

1949

Heliophon
Mastersonic organ

1949
1949

Connsonata

1949

Melochord

19479

B. Helberger
J. Goodell and
E. Swedien
Conn Organ
Company
H. Bode

Bel Organ

c. 1947

Bendix Electronics

Elektronium Pi

1950

Hohner GmbH

Radareed organ
Dereux organ

1950
c. 1950

G. Gubbins
Societe Dereux

1947

B. Cross and
P. Grainger
O. Sala

Voltage-controlled synthesizer, pitch,

waveform, and formant controllers
Electronic oscillators and continuous
automated control
Trautonium with noise generator, ``circuitbreaker'' sequencer, frequency dividers
Rotating pitch wheels
Oscillators designed by E. L. Kent
Later installed at North West German Radio,
Cologne
12 vacuum-tube oscillators, other pitches
obtained by divide-down circuit
Monophonic vacuum-tube oscillator with
divide-down circuitry
Amplied reeds tted with resonators
Electrostatic rotary generators, waveforms
derived from oscillogram photographs

What is a wave? In acoustics it is dened as a disturbance (wavefront) that

propagates continuously through a medium or through space. A wave oscillation moves away from a source and transports no signicant amount of matter
over large distances of propagation.
Optical Wave versus Particle Debate
The waveparticle debate in optics began in the early eighteenth century, when
Isaac Newton, in his Opticks (published in 1704), described light as a stream
of particles, partly because ``it travels in a straight line.'' Through experiments
with color phenomena in glass plates he also recognized the necessity of ascrib-

50

Chapter 2

ing certain wavelike properties to light beams. Newton was careful not to
speculate further, however, and the corpuscular or particle theory of light held
sway for a century (de Broglie 1945; Elmore and Heald 1969).
A competing wave theory began to emerge shortly afterward with the
experiments in reection and refraction of Christian Huygens, who also performed experiments on the wave nature of acoustical signals. The early nineteenth century experiments of Thomas Young reinforced the wave view. Young
observed that a monochromatic beam of light passing through two pinholes
would set up an interference pattern resembling ``waves of water,'' with their
characteristic patterns of reinforcement and cancellation at points of intersection, depending on their phase. Experiments by Augustin Fresnel and others
seemed to conrm this point of view. The theory of electromagnetic energy
proposed by the Scottish physicist James Clerk Maxwell (18311879) described
light as a wave variation in the electromagnetic eld surrounding a charged
particle. The oscillations of the particle caused the variations in this eld.
Physicists resolved the optical waveparticle controversy in the rst two
decades of the twentieth century. This entailed a unied view of matter and
electromagnetic energy as manifestations of the same phenomena, but with
dierent masses. The wave properties of polarization and interference, demonstrated by light, are also exhibited by the atomic constituents of matter, such as
electrons. Conversely, light, in its interaction with matter, behaves as though
composed of many individual units (called photons), which exhibit properties
usually associated with particles, such as energy and momentum.
Acoustical Wave versus Particle Debate
What Atomes make Change
Tis severall Figur'd Atomes that make Change,
When severall Bodies meet as they do range.
For if they sympathise, and do agree,
They joyne together, as one Body bee.
But if they joyne like to a Rabble-rout,
Without all order running in and out;
Then disproportionable things they make,
Because they did not their right places take.
(Margaret Cavendish 1653)

The idea that a continuous tone could be decomposed into smaller quantities of
time emerges from ancient atomistic philosophies. The statement that all matter
is composed of indivisible particles called atoms can be traced to the ancient

51

The History of Microsound

city of Abdera, on the seacoast of Thrace. Here, in the latter part of the fth
century BC, Leucippus and Democritus taught that all matter consists only of
atoms and empty space. These Greek philosophers are the joint founders of
atomic theory. In their opinion, atoms were imperceptible, individual particles
diering only in shape and position. The combination of these particles causes
the world we experience. They speculated that any substance, when divided into
smaller and smaller pieces, would eventually reach a point where it could no
longer be divided. This was the atom.
Another atomist, Epicurus (341270 BC), founded a school in Athens in 306
BC and taught his doctrines to a devoted body of followers. Later, the Roman
Lucretius (55) wrote De Rerum Natura (On the Nature of the Universe) delineating the Epicurean philosophy. In Book II of this text, Lucretius characterized the universe as a fortuitous aggregation of atoms moving in the void.
He insisted that the soul is not a distinct, immaterial entity but a chance combination of atoms that does not survive the body. He further postulated that
earthly phenomena are the result of purely natural causes. In his view, the
world is not directed by divine agency; therefore fear of the supernatural is
without reasonable foundation. Lucretius did not deny the existence of gods,
but he saw them as having no impact upon the aairs of mortals (Cohen 1984,
p. 177).
The atomistic philosophy was comprehensive: both matter and energy (such
as sound) were composed of tiny particles.
Roughness in the voice comes from roughness in its primary particles, and likewise smoothness is begotten of their smoothness. (Lucretius 55, Book IV, verse 524)

At the dawn of early modern science in the seventeenth century, the French
natural philosophers Pierre Gassendi (15921655) and Rene Descartes (1596
1650) revived atomism. Descartes' theory of matter was based on particles and
their motion. Gassendi (1658) based his system on atoms and the void. The
particles within these two systems have various shapes, weights, or other qualities that distinguish them. From 1625 until his death, Gassendi occupied himself with the promulgation of the philosophy of Epicurus.
During the same period, the science of acoustics began to take shape in
western Europe. A conuence of intellectual energy, emanating from Descartes,
Galileo, Beekman, Mersenne, Gassendi, Boyle, and others, gradually forced a
paradigm shift away from the Aristotelian worldview toward a more experimental perspective. It is remarkable how connected was this shift in scientic
thinking to the analysis of musical sound (Coelho 1992). Problems in musical

52

Chapter 2

acoustics motivated experiments that were important to the development of

modern science.
The Dutch scholar Isaac Beekman (15881637) proposed in 1616 a ``corpuscular'' theory of sound. Beekman believed that any vibrating object, such as
a string, cuts the surrounding air into spherical particles of air that the vibrations project in all directions. When these particles impinge on the eardrum, we
perceive sound.
The very same air that is directly touched and aected by a hard thing is violently shocked
and dispersed [by a vibrating object] and scattered particle-wise everywhere, so that the air
itself that had received the impulse strikes our ear, in the way that a candle ame spreads
itself through space and is called light. (Cohen 1984)

In Beekman's theory, the particles emitted by a vibrating string derive their

velocity from the force with which the string hits them. Every particle ies o
on its own, is homogeneous, and represents in its particular shape and size the
properties of the resulting sound. If a particle does not hit the ear, it nally
comes to rest, according to the laws of projectile motion, and is then reintegrated into the surrounding air. Beekman ascribed dierences in timbre to
variations in the size, shape, speed, and density of sound particles. Gassendi
also argued that sound is the result of a stream of particles emitted by a
sounding body. The velocity of sound is the speed of the particles, and frequency is the number of particles emitted per unit time.
Almost two centuries later, in 1808, an English school teacher, John Dalton
(17661844), formulated an atomic theory of matter. Unlike the speculations of
Beekman and Gassendi, Dalton based his theory on experimental evidence
(Kargon 1966). Dalton stated that all matter is composed of extremely small
atoms that cannot be subdivided, created, or destroyed. He further stated
that all atoms of the same element are identical in mass, size, and chemical
and physical properties, and that the properties of the atom of one element,
dier from those of another. What dierentiates elements from one another, of
course, are their constituent particles. Eighty-nine years after Dalton, the rst
elementary particlethe electronwas discovered by another Englishman,
J. J. Thomson (Weinberg 1983).
As the particle theory of matter emerged, however, the particle theory of
sound was opposed by increasing evidence. The idea of sound as a wave
phenomenon grew out of ancient observations of water waves. That sound
may exhibit analogous behavior was emphasized by a number of Greek and
Roman philosophers and engineers, including Chrysippus (c. 240 BC), Vetruvius

53

The History of Microsound

(c. 25 BC), and Boethius (480524). The wave interpretation was also consistent with Aristotle's (384322 BC) statement to the eect that air motion is
generated by a source, ``thrusting forward in like manner the adjoining air, so
that the sound travels unaltered in quality as far as the disturbance of the air
manages to reach.''
By the mid-1600s, evidence had begun to accumulate in favor of the wave
hypothesis. Robert Boyle's classic experiment in 1640 on the sound radiation of
a ticking watch in a partially evacuated glass vessel gave proof that the medium
of air was necessary for the production or transmission of audible sound.
Experiments showed the relation between the frequency of air motion and
the frequency of a vibrating string (Pierce 1994). Galileo Galilei's book Mathematical Discourses Concerning Two New Sciences, published in 1638, contained the clearest statement given until then of frequency equivalence, and, on
the basis of accumulated experimental evidence, Rene Descartes rejected
Beekman's corpuscular theory of sound (Cohen 1984, p. 166).
Marin Mersenne's description in his Harmonie Universelle (1636) of the rst
absolute determination of the frequency of an audible tone (at 84 Hz) implies
that he had already demonstrated that the absolute-frequency ratio of two
vibrating strings, radiating a musical tone and its octave, is as 1 : 2. The perceived harmony (consonance) of two such notes could be explained if the ratio
of the air oscillation frequencies is also 1 : 2, which is consistent with the wave
theory of sound.
Thus, a continuous tone could be decomposed into small time intervals, but
these intervals would correspond to the periods of a waveform, rather than to
the rate of ow of sonic particles.
The analogy with water waves was strengthened by the belief that air motion associated with musical sounds is oscillatory and by the observation that
sound travels with a nite speed. Another matter of common knowledge was
that sound bends around corners, suggesting diraction, also observed in water
waves (gure 2.1). Sound diraction occurs because variations in air pressure
cannot go abruptly to zero after passing the edge of an object. They bend, instead, into a shadow zone in which part of the propagating wave changes direction and loses energy. This is the diracted signal. The degree of diraction
depends on the wavelength (short wavelengths diract less), again conrming
the wave view.
While the atomic theory of matter became the accepted viewpoint in the
nineteenth century, the wave theory of sound took precedence. New particlebased acoustic theories were regarded as oddities (Gardner 1957).

54

Chapter 2

Figure 2.1 Zones of audition with respect to a sound ray and a corner. Listeners in
zone A hear the direct sound and also the sound reected on the wall. Those in zone B
hear a combination of direct, reected, and diracted sound. In zone C they hear a
combination of direct and diracted sound. Listeners in zone D hear only diracted
sound (after Pierce 1994).

Waves versus Particles: a Contemporary Perspective

The wave theory of sound dominated the science of acoustics until 1907, when
Albert Einstein predicted that ultrasonic vibration could occur on the quantum
level of atomic structure, leading to the concept of acoustical quanta or phonons.
Einstein's theory of phonons was nally veried in 1913.
In his own way, the visionary composer Edgard Varese recognized the signicance of this discovery:
Every tone is a complex entity made up of elements ordered in various ways . . . In other
words, every tone is a molecule of music, and as such can be dissociated into component
sonal atoms. . . . [These] may be shown to be but waves of the all-pervading sonal energy
radiating throughout the universe, like the recently discovered cosmic rays which Dr. Milliken calls, interestingly enough, the birth cries of the simple elements: helium, oxygen,
silicon, and iron. (Varese 1940)

The scientic development of acoustical quantum theory in the domain

of audible sounds was left to the physicist Dennis Gabor (1946, 1947, 1952).
Gabor proposed that all sound could be decomposed into a family of functions

55

The History of Microsound

obtained by time and frequency shifts of a single Gaussian particle. Gabor's

pioneering ideas have deeply aected signal processing and sound synthesis.
(See chapters 3 and 6.) Later in this chapter, we present the basic idea of the
Gabor matrix, which divides time and frequency according to a grid.
Today we would say that the wave and particle theories of sound are not
opposed. Rather, they reect complementary points of view. In matter, such as
water, waves move on a macro scale, but water is composed of molecules
moving on a micro scale. Sound can be seen in a similar way, either wavelike or
particle-like, depending upon the scale of measurement, the density of particles,
and the type of operations that we apply to it.

The Modern Concept of Microsound

Fundamental to microsound synthesis is the recognition of the continuum between rhythm (the infrasonic frequencies) and pitch (the audible frequencies).
This idea was central to what the poet and composer Ezra Pound called the
theory of the ``Great Base'' (Pound 1934). In 1910 he wrote:
Rhythm is perhaps the most primal of all things known to us . . . Music is, by further
analysis, pure rhythm; rhythm and nothing else, for the variation of pitch is the variation in
rhythms of the individual notes, and harmony, the blending of these varied rhythms. (Pound
1910, in Schafer 1977)

Pound proposed the Great Base theory in 1927:

You can use your beat as a third or fourth or Nth note in the harmony. To put it another
way; the percussion of the rhythm can enter the harmony exactly as another note would. It
enters usually as a Bassus . . . giving the main form to the sound. It may be convenient to
call these dierent degrees of the scale the megaphonic and microphonic parts of the harmony. Rhythm is nothing but the division of frequency plus an emphasis or phrasing of that
division. (Pound 1927, in Schafer 1977)

In this theory, Pound recognized the rhythmic potential of infrasonic frequencies. The composer Henry Cowell also describes this relationship:
Rhythm and tone, which have been thought to be entirely separate musical fundamentals
. . . are denitely related through overtone ratios. (Cowell 1930)

Later in his book he gives an example:

Assume that we have two melodies in parallel to each other, the rst written in whole notes
and the second in half-notes. If the time for each note were to be indicated by the tapping of

56

Chapter 2
a stick, the taps for the second melody would recur with double the rapidity of those of the
rst. If now the taps were to be increased greatly in rapidity without changing the relative
speed, it will be seen that when the taps for the rst melody reach sixteen to the second,
those for the second melody will be thirty-two to the second. In other words, the vibrations
from the taps of one melody will give the musical tone C, while those of the other will give
the tone C one octave higher. Time has been translated, as it were, into musical tone. Or, as
has been shown above, a parallel can be drawn between the ratio of rhythmical beats and
the ratio of musical tones by virtue of the common mathematical basis of both musical time
and musical tone. The two times, in this view, might be said to be ``in harmony,'' the simplest possible. . . . There is, of course, nothing radical in what is thus far suggested. It is only
the interpretation that is new; but when we extend this principle more widely we begin to
open up new elds of rhythmical expression in music. (Cowell 1930)

Cowell formulated this insight two decades before Karlheinz Stockhausen's

temporal theory, explained later in this chapter.
Temporal Continuity in Perception
Inherent in the concept of microsound is the notion that sounds on the object
time scale can be broken down into a succession of events on a smaller time
scale. This means that the apparently continuous ow of music can be considered as a succession of frames passing by at a rate too fast to be heard as discrete events. This ideal concept of time division is ancient (consider Zeno of
Elea's four paradoxes). It could not be fully exploited by technology until the
modern age.
In the visual domain, the illusion of cinemamotion picturesis made possible by a perceptual phenomenon known as persistence of vision. This enables
a rapid succession of discrete images to fuse into the illusion of a continuum.
Persistence of vision was rst explained scientically by P. M. Roget in 1824
(Read and Welch 1977). W. Fritton demonstrated it with images on the two
sides of a card: one of a bird, the other of a cage. When the card was spun
rapidly, it appeared that the bird was in the cage (de Reydellet 1999).
The auditory analogy to persistence of vision is the phenomenon of tone
fusion induced by the forward masking eect, described in chapter 1.
Throughout the nineteenth century, slow progress was made toward the development of more sophisticated devices for the display of moving images. (See
Read and Welch 1977 for details.) A breakthrough, however, did come in 1834
with W. G. Horner's Zoetrope (originally called the Daedelum). The Zoetrope
took advantage of persistence of vision by rotating a series of images around a

57

The History of Microsound

xed window tted with a viewing lens. Depending on the speed of rotation, the
image appeared to move in fast or slow motion.
After the invention of celluloid lm for photography, the ubiquitous Thomas
Alva Edison created the rst commercial system for motion pictures in 1891.
This consisted of the Kinetograph camera and the Kinetoscope viewing system.
Cinema came into being with the projection of motion pictures onto a large
screen, introduced by the Lumiere brothers in 1895.
In 1889 George Eastman demonstrated a system which synchronized moving
pictures with a phonograph, but the ``talking picture'' with optical soundtrack
did not appear until 1927. An optical sound track, however, is not divided into
frames. It appears as a continuous band running horizontally alongside the
succession of vertical image frames.
In music, automated mechanical instruments had long quantized time into
steps lasting as little as a brief note. But it was impossible for these machines to
operate with precision on the time scale of microsound. Electronics technology
was needed for this, and the modern era of microsound did not dawn until the
acoustic theory and experiments of Dennis Gabor in the 1940s.
The Gabor Matrix
Inherent in the concept of a continuum between rhythm and pitch is the notion
that tones can be considered as a succession of discrete units of acoustic energy.
This leads to the notion of a granular or quantum approach to sound, rst
proposed by the British physicist Dennis Gabor in a trio of brilliant papers.
These papers combined theoretical insights from quantum physics with practical experiments (1946, 1947, 1952). In Gabor's conception, any sound can be
decomposed into a family of functions obtained by time and frequency shifts of
a single Gaussian particle. Another way of saying this is that any sound can
be decomposed into an appropriate combination of thousands of elementary
grains. It is important to emphasize the analytical orientation of Gabor's
theory. He was interested in a general, invertible method for the analysis of
waveforms. As he wrote in 1952:
The orthodox method [of analysis] starts with the assumption that the signal s is a function s(t) of time t. This is a very misleading start. If we take it literally, it means that we
have a rule of constructing an exact value of s(t) to any instant of time t. Actually we are
never in a position to do this. . . . If there is a bandwidth W at our disposal, we cannot mark
time any more exactly than by a time-width of the order 1/W; hence we cannot talk
physically of time elements smaller than 1/W. (Gabor 1952, p. 6)

58

Chapter 2

Gabor took exception to the notion that hearing was well represented by
Fourier analysis of innite signals, a notion derived from Helmholtz (1885). As
he wrote:
Fourier analysis is a timeless description in terms of exactly periodic waves of innite duration. On the other hand it is our most elementary experience that sound has a time pattern as well as a frequency pattern. . . . A mathematical description is wanted which ab ovo
takes account of this duality. (Gabor 1947, p. 591)

Gabor's solution involved the combination of two previously separated

dimensions: frequency and time, and their correlation in two new representations: the mathematical domain of acoustic quanta, and the psychoacoustical
domain of hearing. He formed a mathematical representation for acoustic
quanta by relating a time-domain signal st to a frequency-domain spectrum
S f . He then mapped an energy function from st over an ``eective duration'' Dt into an energy function from S f over an ``eective spectral width''
D f to obtain a characteristic cell or acoustic quantum. Today one refers to
analyses that are limited to a short time frame as windowed analysis (see chapter
6). One way to view the Gabor transform is to see it as a kind of collection of
localized Fourier transforms. As such, it is highly useful for the analysis of
time-varying signals, such as music.
Gabor recognized that any windowed analysis entails an uncertainty relation
between time and frequency resolution. That is, a high resolution in frequency
requires the analysis of a large number of samples. This implies a long time
window. It is possible to pinpoint specic frequencies in an analyzed segment of
samples, but only at the cost of losing track of when exactly they occurred.
Conversely, it is possible to pinpoint the temporal structure of audio events
with great precision, but only at the cost of giving up frequency precision. This
relation is expressed in Gabor's formula:
Dt  D f b 1
For example, if the uncertainty product is 1 and Dt is 10 ms (or 1/100 Hz), then
D f can be no less than 100 Hz. Another way of stating this is: to resolve
frequencies to within a bandwidth of 100 Hz, we need a time window of at least
10 ms.
Time and frequency resolution are bound together. The more precisely we x
one magnitude, the more inexact is the determination of the other.
Gabor's quanta are units of elementary acoustical information. They can be
represented as elementary signals with oscillations at any audible frequency f ,

59

The History of Microsound

modulated by a nite duration envelope (a Gaussian curve). Any audio signal

fed into a Gabor analyzer can be represented in terms of such signals by expanding the information area (time versus frequency) into unit cells and associating
with each cell an amplitude factor (gure 2.2). His formula for sound quanta
was:
gt ea

tt0 2

 e 2pjf0 t

where
Dt p 1=2 =a

and

D f a=p 1=2

The rst part of equation 1 denes the Gaussian envelope, while the second
part denes the complex sinusoidal function (frequency plus initial phase)
within each quantum.
The geometry of the acoustic quantum Dt D f depends on the parameter a,
where the greater the value of a, the greater the time resolution at the expense
of the frequency resolution. (For example, if a 1:0, then Dt 1:77245, and
D f 0:56419. Setting the time scale to milliseconds, this corresponds to a time
window of 1.77245 ms, and a frequency window of 564.19 Hz. For a 2:0, Dt
would be 0.88 ms and D f would be 1128.38 Hz.) The extreme limiting cases of
the Gabor series expansion are a time series (where Dt is the delta function d),
and the Fourier series (where Dt y).
Gabor proposed that a quantum of sound was a concept of signicance to
the theory of hearing, since human hearing is not continuous and innite in
resolution. Hearing is governed by quanta of dierence thresholds in frequency, time, and amplitude (see also Whiteld 1978). Within a short time window (between 10 and 21 ms), he reasoned, the ear can register only one distinct
sensation, that is, only one event at a specic frequency and amplitude.
Gabor gave an iterative approximation method to calculate the matrix. By
1966 Helstrom showed how Gabor's analysis/resynthesis approximation could
be recast into an exact identity by turning the elementary signals into orthogonal functions. Bacry, Grossman, and Zak (1975) and Bastiaans (1980, 1985)
veried this hypothesis. They developed analytic methods for calculating the
matrix and resynthesizing the signal.
A similar time-frequency lattice of functions was also proposed in 1932 in a
dierent context by the mathematician John von Neumann. It subsequently
became known as the von Neumann lattice and lived a parallel life among
quantum physicists (Feichtinger and Strohmer 1998).

60

Chapter 2

61

The History of Microsound

Electro-optical and Electromechanical Sound Granulation

Gabor was also an inventor, and indeed, he won the Nobel Prize for the invention of holography. In the mid-1940s, he constructed a sound granulator
based on a sprocketed optical recording system adapted from a 16 mm lm
projector (Gabor 1946). He used this ``Kinematical Frequency Convertor'' to
make pitch-time changing experimentschanging the pitch of a sound without
changing its duration, and vice versa.
Working with Pierre Schaeer, Jacques Poullin constructed another spinning-head device, dubbed the Phonogene, in the early 1950s (Schaeer 1977,
pp. 4179, 4278; Moles 1960). (See also Fairbanks, Everitt, and Jaeger 1954
for a description of a similar invention.) Later, a German company, Springer,
made a machine based on similar principles, using the medium of magnetic tape
and several spinning playback heads (Morawaska-Bungler 1988; Schaeer
1977, pp. 4278). This device, called the Zeitregler or Tempophon, processed
speech sounds in Herbert Eimert's 1963 electronic music composition Epitaph
fur Aikichi Kuboyama (recorded on Wergo 60014).
The basic principle of these machines is time-granulation of recorded sounds.
In an electromechanical pitch-time changer, a rotating head (the sampling
head) spins across a recording (on lm or tape) of a sound. The sampling head
spins in the same direction that the tape is moving. Because the head only
contacts the tape for a short period, the eect is that of sampling the sound on
the tape at regular intervals. Each of these sampled segments is a grain of
sound.
In Gabor's system, the grains were reassembled into a continuous stream on
another recorder. When this second recording played back, the result was a
more-or-less continuous signal but with a dierent time base. For example,
shrinking the duration of the original signal was achieved by slowing down the
rotation speed of the sampling head. This meant that the resampled recording
contained a joined sequence of grains that were formerly separated. For time
expansion, the rotating head spun quickly, sampling multiple copies (clones) of
the original signal. When these samples were played back as a continuous
signal, the eect of the multiple copies was to stretch out the duration of the
Figure 2.2 The Gabor matrix. The top image indicates the energy levels numerically.
The middle image indicates the energy levels graphically. The lower image shows how
the cells of the Gabor matrix (bounded by Dn, where n is frequency, and Dt, where t is
time) can be mapped into a sonogram.

62

Chapter 2

resampled version. The local frequency content of the original signal and in
particular of the pitch, is preserved in the resampled version.
To eect a change in pitch without changing the duration of a sound, one
need only to change the playback rate of the original and use the timescale modication just described to adjust its duration. For example, to shift the pitch
up an octave, play back the original at double speed and use time-granulation
to double the duration of the resampled version. This restores the duration
to its original length. Chapter 5 looks at sound granulation using digital
technology.
Meyer-Eppler
The acoustician Werner Meyer-Eppler was one of the founders of the
West Deutscher Rundfunk (WDR) studio for electronic music in Cologne
(Morawska-Bungler 1988). He was well aware of the signicance of Gabor's
research. In an historic lecture entitled Das Klangfarbenproblem in der elektronischen Musik (``The problem of timbre in electronic music'') delivered in
August 1950 at the Internationale Ferienkurse fur Neue Musik in Darmstadt,
Meyer-Eppler described the Gabor matrix for analyzing sounds into acoustic
quanta (Ungeheuer 1992). He also presented examples of Oskar Fischinger's
animated lms with their optical images of waveforms as the ``scores of the
future.'' In his later lecture Metamorphose der Klangelemente, presented in 1955
at among other places, Gravesano, Switzerland at the studio of Hermann
Scherchen, Meyer-Eppler described the Gabor matrix as a kind of score that
could be composed with a ``Mosaiktechnik.'' In his textbook, Meyer-Eppler
(1959) described the Gabor matrix in the context of measuring the information
content of audio signals. He dened the ``maximum structure content'' of a
signal as a physical measurement
K 2W T
where W is the bandwidth in Hertz and T is the signal duration. Thus for a
signal with a full bandwidth of 20 kHz and a duration of 10 seconds, the
maximum structure content is 2  20000  10 400;000, which isby the sampling theoremthe number of samples needed to record it. He recognized that
aural perception was limited in its time resolution, and estimated that the
lower boundary on perception of parameter dierences was of the order of 15 ms,
about 1/66th of a second.

63

The History of Microsound

The concept of time-segmentation was central to his notion of systematic

sound transformation (Meyer-Eppler 1960). For example, he described experiments with speech in which grains from one word could be interpolated into
another to change its sense.
Moles
The physicist Abraham Moles (1960, 1968) was interested in applying Shannon's information theory to aesthetic problems, particularly in new music
(Galante and Sani 2000). Pierre Schaeer hired him to work at the Groupe de
Recherches Musicale (GRM). Signicantly, this coincided with Iannis Xenakis's
residencies in the GRM studios (Orcalli 1993). Moles had read Meyer-Eppler's
book. He sought a way to segment sound objects into small units for the purpose of measuring their information content. Following the Gabor matrix, he
set up a three-dimensional space bounded by quanta in frequency, loudness,
and time. He described this segmentation as follows:
We know that the receptor, the ear, divides these two dimensions [pitch and loudness]
into quanta. Thus each sonic element may be represented by an elementary square. A
pure sinusoidal sound, without any harmonics, would be represented by just one of these
squares. . . . Because thresholds quantize the continua of pitch and loudness, the repertoire
is limited to some 340,000 elements. Physically, these elements are smaller and denser toward the center of the sonic domain, where the ear is more acute. . . . In most cases each
symbol [in a sonic message] is a combination of elements, that is, of a certain number of
these squares. (Moles 1968)

Wiener
The MIT mathematician Norbert Wiener (the founder of cybernetics) was well
aware of Gabor's theory of acoustic quanta, just as Gabor was well aware of
Wiener's work. In 1951, Gabor was invited to present his acoustical quantum
theory in a series of lectures at MIT (Gabor 1952).
Like Gabor, Wiener rejected the view (expounded by Leibniz in the eighteenth century) that time, space, and matter are innitely subdivisible or continuous. He supported Planck's quantum theory principle of discontinuity in
light and in matter. Wiener noted that Newton's model of deterministic physics
was being replaced by Gibbsian statistical mechanicsa ``qualied indeterminism.'' And like Gabor, he was skeptical of Fourier analysis as the best representation for music.

64

Chapter 2
The frequency and timing of a note interact in a complicated manner. To start and stop a
note involves an alteration of its frequency content which may be small but very real. A
note lasting only a nite time is to be analyzed as a band of simple harmonic motions, no
one of which can be taken as the only simple harmonic motion present. The considerations
are not only theoretically important but correspond to a real limitation of what a musician
can do. You can't play a jig in the lowest register of an organ. If you take a note oscillating
at sixteen cycles per second and continue it only for one twentieth of a second, what you get
is a single push of air without any noticeable periodic character. Just as in quantum theory,
there is in music a dierence of behavior between those things belonging to small intervals
of time and what we accept on the normal scale of every day. (Wiener 1964a, 1964b)

Going further, Wiener stressed the importance of recognizing the time scale of a
model of measurement:
The laws of physics are like music notationthings that are real and important provided
that we do not take them too seriously and push the time scale down below a certain
level. (Wiener 1964a, 1964b)

Theory and Experiments of Xenakis

Iannis Xenakis leaves many legacies. Besides his remarkably inventive compositional output, he was one of the great free-thinkers in the history of music
theory. He expanded the mathematical foundations of music in all its dimensions: pitch and scale, rhythm, timbre, sound synthesis, composition strategy,
and form. Unlike most musicians, Xenakis constantly kept aware of developments in science and engineering. This knowledge fed his musical theories.
A fascination for statistical and particulated sound textures is apparent in
his rst orchestral composition Metastasis (1954). This interest carries through
to his 1958 electroacoustic composition Concret PH, realized at the Groupe de
Recherches Musicale (GRM) studios in Paris and premiered at the Philips
Pavilion at the Brussels World's Fair. To create the granular texture for
this work, he mixed recordings of burning wood-embers, cut into one-second
fragments (Solomos 1997). These crackling sound mixtures were manipulated
slightly in the studio of the GRM. Describing this work, he said:
Start with a sound made up of many particles, then see how you can make it change imperceptibly, growing and developing, until an entirely new sound results. . . . This was in
deance of the usual manner of working with concrete sounds. Most of the musique concrete which had been produced up to the time of Concret PH is full of many abrupt changes
and juxtaposed sections without transitions. This happened because the original recorded
sounds used by the composers consisted of a block of one kind of sound, then a block of
another, and did not extend beyond this. I seek extremely rich sounds (many high over-

65

The History of Microsound

tones) that have a long duration, yet with much internal change and variety. Also, I explore
the realm of extremely faint sounds highly amplied. There is usually no electronic
alteration of the original sound, since an operation such as ltering diminishes the
richness. (Xenakis program notes, Nonesuch recording H-71246)

The main techniques employed in Concret PH are the splicing of numerous

bits of magnetic tape, tape speed change, and mixing to obtain varying
densities. This work was not the result of mathematical operations, but was
approached by the composer intuitively, in the manner of sound sculpture.
Organization of Analogique B
A year after Concret PH, Xenakis tried working more systematically with
sound grains. He proposed the hypothesis that every sound could be understood as the assembly of a number of elementary particles. In Analogique B,
completed in 1959 and premiered in Gravesano later that year, he drew from
Moles's research at the GRM, and from Meyer-Eppler's (1959) book on information theory, which describes the Gabor matrix.
Analogique B consists of granular sounds produced by recording sine tones
emitted by analog tone generators onto analog tape, and then cutting the tones
into fragments. This brief composition (2 minute 25-sec) was meant to be
played after Analogique A, a stochastically composed score for two violins, two
cellos, and two contrabasses. Chapter 3 of Formalized Music (Xenakis 1971,
1992) describes how he organized the synthetic grains in Analogique B (1959,
Editions Salabert). See also Di Scipio (1995, 1997a, 1998), and Harley (forthcoming). The graphic score of Analogique B appears in the front matter of
Formalized Music.
Analogique B was designed by scattering grains onto time-grids, called
screens by Xenakis. The screens represented elementary sonic quanta in three
dimensions: dierence thresholds in frequency, amplitude, and time. He coined
the term ``grains of sound'' (Xenakis 1960), and was the rst musician to explicate a compositional theory for sound grains. He proposed the following lemma:
All sound, even continuous musical variation, is conceived as an assemblage of a large
number of elementary sounds adequately disposed in time. In the attack, body, and decline
of a complex sound, thousands of pure sounds appear in a more or less short interval of
time Dt. (Xenakis 1992, p. 43)

Synchronized with the advancing time interval Dt, the screens are snapshots
of sound bounded by frequency and amplitude grids, each screen subdivided

66

Chapter 2

into elementary squares of sonic energy, according to the resolution of the

Gabor matrix. Xenakis goes on to specify a mesostructural unit describing a
sequence of screens, which he calls a book. A book of screens could constitute
the entirety of a complex sounda cloud of points in evolution.
How should the grains be distributed on the screen? For a given density,
Xenakis turned to probability theory for the answer. He proposed an exponential distribution to determine the duration of the strands, which were
divided up between a succession of screens. This formula is as follows:
Px cecx dx
where c is the mean duration and x is the time axis. This equation describes the
probability P that an event of duration xi between x and x dx will occur. (See
Lorrain 1980.) For the frequency, amplitude, and density parameters of the
grains he proposed the linear distribution rule:
Pg 2=a1 g=a dg
which gives the probability that a segment (or interval) of length a will have a
length included within g and g dg, for 0 a g a a. Such a formula favors
smaller intervals. Setting a 10, for example, the probability of the small interval 2 is 0.16, while for the larger interval 9 it is 0.02.
Xenakis observed how sound particles could be viewed as short vectors
within a three-dimensional space bounded by frequency, amplitude, and time.
The Gabor grain, with its constant frequency and amplitude, is a special case of
this view. It could also be possible to create grains that were short glissandi, for
example (Xenakis 1960, p. 100). (I have implemented this idea in the technique
of glisson synthesis, described in chapter 4.)
After dening the framework, Xenakis proposed a set of transformations
that could be applied to the screens in order to create new screens. New screens
could be generated, for example, by taking the logical intersection of two
screens. He also proposed other Boolean operations, such as set union, complement, and dierence.
His next theoretical elaboration scrutinized the ataxy or degree of order versus disorder in a succession of screens. Maximum disorder, for example, would
correspond to extreme changes in the distribution of frequency and amplitude
energy, creating a sonic eect akin to white noise. Perfect order would correspond to a solitary sine wave extending across multiple screens.
The ow of ataxy could be regulated via a matrix of transition probabilities,
otherwise known as a Markov chain (Roads 1996). A Markov chain lets one

67

The History of Microsound

encode transitions from order to disorder as a set of weighted probabilities,

where the probability at time t depends on the history at times t 1, t 2, etc.
Here is a simple transition matrix for a rst-order Markov chain. It is called
rst-order because it only looks back one step.
A
B
C

A
0.1
0.33
0.8

B
0.9
0.33
0.1

C
0
0.34
0.1

This matrix indicates the probabilities of three outcomes A, B, C, given three

possible previous states A, B, C. For a given previous state, indicated by the
columns, we read across the rows to determine the probabilities for the next
state. The probabilities in a row add up to 1. Given the previous state A, for
example, we see that A has a 0.1 chance of occurring again, while B has a 0.9
chance, and C has a probability of 0. Thus C will never follow A, and it is very
likely that B will follow A. In granular synthesis, the states could be grain frequencies, amplitudes, or densities.
At the same time Xenakis recognized that ataxy as a compositional principle
was incomplete. For example, certain congurations of grains on the plane of
frequency-versus-amplitude engage the listener, while others do not, even if both
measure the same in terms of ataxy.
Problems with a Constant Microtime Grid
In Xenakis's theory of screens, the assumption is that the frame rate and the
grain duration are constant. The frame rate would determine the smallest grain
size. The idea that all grains have the same duration is aesthetically limiting.
Experiments (described in chapter 3) show that grain size is one of the most
important time-varying parameters of granular synthesis.
The constant microtemporal grid or frame rate Dt of the screens poses technical problems. The main problem being that such a frame rate will tend to
cause audible artefacts (a constant modulation or comb ltering eect, depending on the precise rate) unless countermeasures are taken. Consider a single
stream of grains, one following the next, each lasting 40 ms, each with a Gaussian envelope. The attack of each grain takes 510 ms, as does its decay. This
creates a regular amplitude modulation eect, producing sidebands around the
carrier frequency.

68

Chapter 2

If one were to implement Xenakis's screens, one would want to modify the
theory to allow Dt to be less than the grain duration. This measure would allow
grain attacks and decays to overlap, thereby smoothing over the perception of
the frame rate. Similar problems of frame rate and overlap are well known
in windowed analysis-resynthesis techniques such as the short-time Fourier
transform (STFT). Frame-based representations are fragile, since any transformation of the frames that perturbs the perfect summation criteria at the boundaries of each frame leads to audible distortions (see chapter 6).
We face the necessity for a synchronous frame rate in any real-time implementation of granular synthesis. Ideally, however, this frame rate should operate at a speed as close as possible to the audio sampling rate.
Analog Impulse Generators
The most important sound particle of the 1950s, apart from those identied
in Xenakis's experiments, was the analog impulse. An impulse is a discrete
amplitude-time uctuation, producing a sound that we hear as a click.
Although the impulse is ideally a narrow rectangular shape, in practice it may
be band-limited or have a ramped attack and decay. An impulse generator
emits a succession of impulses at a specied frequency. Impulse generators
serve many functions in a laboratory, such as providing a source for testing
the impulse response (IR) of a circuit or system. The IR is an important system
measurement (see chapter 5).
The common analog circuit for impulse and square wave generation is the
multivibrator (gure 2.3). Multivibrators can be built using many electronic
technologies: vacuum tubes, transistors, operational ampliers, or logic gates.
Although sometimes referred to as an oscillator, a multivibrator is actually an
automatic switch that moves rapidly from one condition to another, producing
a voltage impulse which can be positive, negative, or a combination of the two.
The multivibrator circuit has the advantage that it is easily tuned to a specic
frequency and duty cycle by adjusting a few circuit elementseither resistance
or capacitance values (Douglas 1957).
The multivibrator was used in electronic music instruments as early as 1928,
in Rene Bertrand's Dynaphone (Rhea 1972). Musicians appropriated laboratory impulse generators in the electronic music studios of the 1950s. Karlheinz
Stockhausen and Gottfried Michael Koenig worked extensively with impulse
generators at the Cologne studio. As Koenig observed:

69

The History of Microsound

Figure 2.3 A multivibrator circuit, after Douglas (1957). Suppose that when switching
on, a small positive voltage appears at V1. This increases the anode current of V1, and in
so doing increases the anode potential of V1, which is communicated to the grid of V2.
As the voltage of V2 falls, so will the anode current of V2, causing a rise in the anode
potential of V1, making it more positive. The process continues until it reaches the cuto
voltage of the vacuum tube. The circuit stays in this condition while the grid of V2 leaks
away at a rate depending on the time constant of C1 and R1. As soon as the anode potential of V2 reaches a point where anode current can ow again, the anode potential of
V2 will fall again since the current is increasing, which drives the grid of V1 negative.
The whole process is continued in the opposite direction until V1 is cut o, and so on
continuously. If C1 C2 and R1 R2 the waveform is symmetrical (square) and has
only odd harmonics.
[The pure impulse] has no duration, like sinus and noise, but represents a brief energy
impetus, comparable to a leaping spark. Consequently it has neither pitch nor timbre. But
it encounters an object and sets it vibrating; as pitch, noise, or timbre of the object which
has been impelled. (Koenig 1959)

Stockhausen's great 1960 composition Kontakte, realized with assistance

from Koenig (Supper 1997), is based entirely on ltered impulses. Figure 2.4
shows the patch interconnections used in its realization, all of which begin with
impulse generation. The technique of recirculating tape feedback loops, seen in
many of the patches, was developed in 1951 by Werner Meyer-Eppler, Stockhausen's teacher (Ungeheuer 1992, p. 121).
Kaegi (1967) describes applications of impulse generators in electronic music.
Chapter 4 presents applications of impulse generators (trainlets and pulsars) in
digital synthesis.

70

Chapter 2

Figure 2.4 Synthesis patches used in the creation of Kontakte by Stockhausen. The
components include impulse generators (IG), preampliers (P), analog tape recorders,
bandpass lters ( f ), and plate reverberators (R). Feedback loops appears as arrows
pointing backwards. (a) Simple impulse generation and recording. (b) Impulse generation with preamplication, ltering, and tape feedback. (c) Impulse generation with
preamplication and ltered feedback. (d) Impulse generation with preamplication,
and multiband ltering. (e) Impulse generation with preamplication, multiband ltering, and tape feedback. (f ) A four-stage process involving (f1) Impulse generation, preamplication, ltering, and recording. (f2) Reverberation with feedback and recording.
(f3) Tape feedback and recording. (f4) Reverberation, ltering, preamplication, and
recording.

71

The History of Microsound

Stockhausen's Temporal Theory

In the 1950s, the Cologne school of electronic music emerged from the studios
of the West Deutscher Rundfunk (WDR). It posited that a composition could
be assembled out of a small number of elementary signals, such as sine waves,
impulses, and ltered noise. In the rst issue of the inuential journal die
Reihe, published in Cologne in 1955, this concept is championed by Karlheinz
Stockhausen, Gottfried Michael Koenig, Herbert Eimert, Karel Goeyvaerts,
and Paul Gredinger. In the same issue, Pierre Boulez, who later turned against
purely electronic music, could not resist this tide of enthusiasm for electronic
sound materials:
Electronic music compels us to assemble each note as we require it.

(Boulez 1955)

Ernst Krenek, one of the rst composers to own an electronic music synthesizer, seemed to anticipate the notion of sound quanta when he mused:
The next step might be the splitting of the atom (that is, the sine tone).

(Krenek 1955)

Could he, perhaps, have been thinking of the Gabor quanta?

The Cologne composers were optimistic about synthesizing interesting forms
of musical sound through combinations of sine waves (Eimert 1955; Goeyvaerts
1955). Many fascinating sounds, however, have transient uctuations that are
not well modelled with sinusoids. Stockhausen's sinusoidal Electronic Etude I
of 1953 was important in its day, but now sounds like a sterile exercise. The
initial enthusiasm for composition using only sine waves soon evaporated. By
the time of Gesang der Junglinge (1956), Stockhausen had moved ahead to identify eleven sound sources:
1. Sine tones
2. Sine tones in which the frequency modulates periodically
3. Sine tones in which the frequency modulates statistically
4. Sine tones in which the amplitude modulates periodically
5. Sine tones in which the amplitude modulates statistically
6. Periodic combinations of both frequency and amplitude modulation
7. Statistical combinations of both frequency and amplitude modulation
8. Colored noise with constant density
9. Colored noise with statistically varying density

72

Chapter 2

10. Periodic sequences of ltered clicks (impulses)

11. Statistical sequences of ltered clicks
He also formulated a theory of the continuum between rhythm and pitch,
that is, between infrasonic frequencies and the audible frequencies of impulses.
If the rate of beat is gradually increased beyond the time constant of the lter and the
limits beyond which the ear can no longer dierentiate, what started as a rhythmically repeated note becomes continuous. . . . We see a continuous transition between what might be
called durational intervals which are characterized as rhythmic intervals and durational
intervals characterized as pitch levels. (Stockhausen 1955)

Stockhausen used an impulse generator to create a regular pulse train. To

the output of this generator he applied a narrow bandpass lter, giving each
pulsation a sharp resonance. If the band was narrow enough, the impulse
resonated around a specic pitch or interval. If the pulse train was irregular,
the infrasonic frequencies generated ametrical rhythms. By transposing these
rhythms into the audible frequency range, Stockhausen could build unpitched
noises from aperiodic sequences of impulses.
Further development of this approach led to a pair of landmark papers on
the composition of microsound, discussed in the next two sections.
How Time Passes
Stockhausen's text ``. . . . . How time passes . . . . .'' was one of many controversial pronouncements made by the composer (Stockhausen 1957). Written over
two months in 1956, when he was 28, and published immediately, it is a raw
outpouring of intellectual reection. The text clearly could have been improved
by critical editing: goals are not stated at the outset, the text unfolds as one long
rambling discourse, and the composer poses problems oering diering solutions. As his exposition proceeds, new criteria are introduced making previous solutions inadequate, so the argument is constantly shifting. Despite these
aws, ``. . . . . How time passes . . . . .'' stands as an ingeniously detailed analysis
of certain relationships between dierent musical time scales, summarized here.
Stockhausen's article has been criticized for its non-standard acoustical
terminology, found in both the original German as well as the English translation by Cornelius Cardew. (The republication of the German edition, in Stockhausen (1963), contains a statement acknowledging the use of nonstandard
terminology.) For example, instead of the common term ``period,'' denoting

73

The History of Microsound

the time interval spanning one cycle of a waveform, Stockhausen uses the term
``phase'' (Phasen), referring not to a ``fundamental period'' but to a ``fundamental phase.'' He substitutes the term ``formant'' for ``harmonic,'' so a harmonic spectrum built up of a fundamental and integer-multiple frequencies is
called a ``formant spectrum.'' He applies the term ``eld'' (Feld ) to denote an
uncertainty region (or band) around a time interval or a central frequency. As
long as one understands these substitutions of terms, however, one can follow
Stockhausen's arguments. In the representation of his article below, I replace
Stockhausen's neologisms with standard acoustical terminology. Page numbers
refer to the English translation.
The most important insight of ``. . . . . How time passes . . . . .'' is a unied view
of the relationship between the various time scales of musical structure. Stockhausen begins by noting the generality of the concept of period, an interval
between two cycles. Period appears in both rhythm (from 6 sec to 1/16th of a
sec) and pitch (from about 1/16th sec to about 1/3200th sec). The key here is
that pitch and rhythm can be considered as one and the same phenomenon,
diering only in their respective time scales. Taking this argument deeper into
the microtemporal domain, the tone color or steady-state spectrum of a note
can also be seen as a manifestation of microrhythm over a fundamental frequency. This point of view can also be applied in the macrotemporal domain.
Thus, an entire composition can be viewed as one time spectrum of a fundamental duration. (As noted earlier, this idea was proposed by Ezra Pound in the
1920s, and by Henry Cowell in 1930.)
The bulk of Stockhausen's text applies this viewpoint to a problem spawned
by serial composition theory; that of creating a scale of twelve durations corresponding to the chromatic scale of pitches in the twelve-tone system. The
problem is exacerbated by Stockhausen's desire to notate the result for performance on traditional instruments. Later, after the composer has developed a
method for generating some of the most arcane rhythmic notation ever devised
(see, for example, the scores of Zeitmasse or the Klavierstucken), he turns to the
diculties of indeterminate notation. Let us now look in more detail at these
arguments.
Stockhausen begins by observing a contradiction in twelve-tone composition theory, which rigorously organizes pitch but not, in any systematic way,
rhythm. Since pitch and rhythm can both be considered as dimensions of time,
Stockhausen proposes that they should both be organized using twelve-element
scales. Constructing a scale of durations that makes sense logically and makes
sense perceptually, however, is not simple. Stockhausen presents several strat-

74

Chapter 2

egies. The rst, adopted by unnamed composers, builds a twelve-element

scale by multiplying a small unit, for example, 1  6; 2  6; 3  6; . . . 12  6.
Clearly, if serial selection order is maintained (i.e., if an element can be chosen
a second time only after all other elements in the series have been chosen), the
long durations will overshadow the short durations. Why is this? The total duration of a twelve-element series is 78  6. If we add up the time taken by the
rst four members of the series, their total duration is 10  6. This, however, is
only about 13% of the total duration of the series. At the other extreme, the
last four notes 9  6; 10  6; 11  6; 12  6, take up 42  6, or more than 53%
of the total duration of the series. Nonetheless, this scheme has been used by
(unnamed) composers, with the obvious added constraint that the tempo be
kept constant in order for the duration series to be perceptible.
To inject this scheme with more exibility composers superimposed series on
top of one another. This led to irregular rhythms that could only be perceived
through what Stockhausen calls statistical form-criteria. A corresponding procedure was adopted with pitch, when ``ocks of notes'' sounding in a short time
period, blurred the pitch structure. Stockhausen was uncomfortable with this
forced agreement between pitch and duration series.
This discomfort led to another strategy for constructing a duration scale.
By subdividing a whole into a series of fractional intervals of the form: 1; 1=2;
1=3; 1=4; . . . 1=12. Stockhausen observed that the rhythmic tuplet series (duplet, triplet, quadruplet, . . .) could be considered as analogous to the harmonic
series (of pitches). An eighth-note triplet, for example, might correspond to
the third harmonic duration of a quarter note. One can speak of a fundamental rhythm with harmonically-related rhythms superimposed. A periodic
tone can be seen as having a rhythmic microstructure, with waveform peaks
corresponding to its internal rhythmic intervals. The dierence between meter
and rhythm corresponds to the distinction on the microtime scale between
fundamental tone and tone color, and one could just as well speak of ``harmonic rhythm'' (my term) as of tone color.
For Stockhausen, the next stage was to relate this harmonic duration scale to
the equal-tempered pitch scale according to the principles of twelve-tone serial
composition. This, however, revealed a new problem: the harmonic scale and
the chromatic scale have little in common. So using a harmonic scale for duration and a chromatic scale for pitch countermands the search for a single time
scale for both domains, the point of which is to unify duration and pitch
through a single set of serial intervallic operations.

75

The History of Microsound

So Stockhausen took on the task of constructing a tempered chromatic scale

of durations corresponding to the equal-tempered scale. He divided the ratio
2 : 1 into 12 equal intervals, according to the logarithmic relationship of the
12th root of 2. These 12 do not translate directly into the common symbols of
music notation, an essential requirement for Stockhausen. He adopted a notational solution wherein all time durations would be represented by the same
symbol, a whole note, for example, adjusted by a tempo indication. So to create
a scale from 5 60 M.M. to 5 120, the note's values are set successively to
tempi of 60, 63.6, 67.4, 71.4, 75.6, 80.1, 84.9, 89.9, 95.2, 100.9, 106.9, 113.3,
120. The interval between each of these numbers corresponds to the standard
6% minor second dierence of the equal-tempered scale. The nal value corresponds to 7 60 M.M., and the scale can be transposed easily. In this scheme,
pitch and time are equal under the regime of the twelve-tone system.
In a sleight-of-hand, Stockhausen then translates an arbitrary duration series
from the equal-tempered notation just described into a list of harmonic proportions, without explaining the procedure. Presumably he multiplies tempo
and note value, rounds o the result to integers, and compares subsequent
integers in the series to obtain a list of harmonic (integer) proportions. This is
merely an exercise, for he then returns to insisting on equal-tempered notation,
which, he observes, remains precise and will not be aected by transposition.
His argument veers o course at this point. While insisting on consistently
equal-tempered proportions for pitch and duration, Stockhausen switches to
harmonic proportions for the organization of higher-level groups of rhythmic
phrases. This is clearly inconsistent with his previous argument.
With this harmonically related hierarchy as his premise, Stockhausen proposes that other serial operations be used to create elaborate networks of harmonic proportions. By deleting some notes and tying others, for example, ever
more complicated relationships can arise. Such transformations lend a sense
of aperiodicity to the rhythmic structure, which Stockhausen compares to the
aperiodic microstructure of noise.
Addressing a practicality of performance, Stockhausen writes that if certain
rhythms cannot be realized by instrumentalists synchronized by a single conductor, then the instrumentalists can be divided into groups and synchronized by separate conductors (a procedure employed in his, Gruppen for three
orchestras). He also confronts the problem for performers of accurately playing
the complicated rhythms generated by the manipulations described above. Notational ambiguity adds to the diculty, since one and the same rhythmic for-

76

Chapter 2

mula may be written in dierent ways, some harder than others for performers
to realize. Stockhausen tries to allow for such imprecisions in his theory by
assigning an ``uncertainty band'' (time-eld ) to the dierent notations. These
time-elds could be derived by recording expert instrumentalists playing dierent gures while measuring the precision of their interpretation. (Obviously it
would be impractical if every rhythmic formula in a composition had to be
notated in multiple ways and tested in such a way.) Stockhausen then proposes
that one could serialize the degree of inaccuracy of performance (!).
The degree of notational complexity and the exactness of performance
are inversely related. If metronome markings are changing from measure to
measure, the uncertainty factor increases. Multiple simultaneous tempi and
broad uncertainty bands lead to general confusion. At this point, Stockhausen
switches the direction of his argument to the organization of statistical groups.
All parameters of time can be turned into ranges, for example, leaving it to the
performers to select from within a specied range. Stockhausen points out that
John Cage, in his proportional notation was not interested in proportional
relationships, depending, as they do, on memories of the past. In Cage's compositions, temporal events are not intentionally linked to the past; always one is
in the present.
Stockhausen prefers a system in which determinacy and indeterminacy stand
at opposite poles of a continuum. He seeks a way to notate structural indeterminacy in a determinate way. This involves ``time smearing'' the music by
interpolating grace notes and articulations (staccato, legato, etc.) that ``fade''
into or out of a central note. Indeterminate notation can also be extended to
meso and macrostructure.
The structure of a piece is presented not as a sequence of development in time but as a
directionless time-eld . . . The groups are irregularly distributed on paper and the general
instructions are: Play any group, selected at random. . . . (p. 36)

The important principle in this gamut between determinacy and indeterminacy is the interplay between the rational counting of time and the ``agitation of
time'' by an instrumentalist. The score is no longer the reference for time.
Instead of mechanically quantifying durations that conict with the regularity of metronomic time, [the performer] now measures sensory quanta; he feels, discovers the time of
the sounds; he lets them take their time. (pp. 378)

The domain of pitch can also be notated aleatorically, and the gamut between pitch and noise can be turned into a compositional parameter.

77

The History of Microsound

To fully realize the pitch-noise continuum, he argues, a new keyboard instrument could be built in which a certain key-pressure produces a constant
repetition of waveform periods (a continuous pitched tone), but a stronger
pressure causes aleatoric modulation leading into noise. This ``ideal instrument'' would be able to move from duration to pitch, from tone to noise, and
also be able to alter the timbre and amplitude of the oscillations. Several instruments playing together would be able to realize all of the riches of Stockhausen's temporal theory. After lamenting on how long one might have to wait
for such an instrument, Stockhausen nishes his article by asserting:
It does not seem very fruitful to founder on a contradiction between, on the one hand, a
material that has become uselessinstruments that have become uselessand, on the
other, our compositional conception.

He leaves this thought hanging as the article ends.

It is evident from his voluminous output of pieces for traditional instruments
since 1957 that Stockhausen did not want to wait for such instruments. For
obvious social and economic reasons he chose to stay within traditional
instrumental practice, albeit with many creative extensions. The technology
of synthesis has greatly advanced since 1957. Stockhausen's ideal instrument
is realizable today using real-time sound synthesis and modern controllers
although, sadly, Stockhausen himself has all but ignored such developments.
The Creatovox, described in chapter 5, is one such solution.
The Unity of Musical Time
Stockhausen based his article ``The unity of musical time'' (Stockhausen 1962) on
a radio broadcast he gave in 1961. This concise text summarizes the basic principles of Stockhausen's integrated approach to electronic music composition.
[In working with an impulse generator], one must proceed from a basic concept of a single
unied musical time; and the dierent perceptual categories such as color, harmony and
melody, meter and rhythm, dynamics, and form must be regarded as corresponding to the
dierent components of this unied time. (p. 217)

Stockhausen distinguishes twenty-one time octaves spanning the durations

from 1/16th of a second to 3600 seconds (one hour), which constitute the range
of perceivable events in a music composition. Finally, he describes the procedure by which he crossed over the boundaries of rhythm and pitch in his composition Kontakte, with specic reference to the pivotal section between 16 : 56
and 18 : 26.

78

Chapter 2

Assessment of Stockhausen's Temporal Theory

The important thing . . . is that tone-color is the result of time structure.
1957, p. 19)

(Stockhausen

When it was published, ``. . . . . How time passes . . . . .'' presented a new viewpoint on musical time. This viewpoint is more familiar to us now, yet we can
still appreciate the depth of Stockhausen's insight. Few articles on music from
the 1950s ring with such resonance today.
The quest for a scale of durations for serial composition is no longer a compelling musical problem. Even in the heyday of serial composition, Stockhausen's solution never entered the repertory of common practice. The most
prominent American exponent of serial techniques, Milton Babbitt (1962), explicitly rejected the idea of constructing a duration scale from multiples of an
elementary unit. For Babbitt, the temporal order of the pitches in the row was
more important than the actual durations of the notes. Thus he reduced the
problem of serial organization of time to the organization of the instants at
which notes started, which he called the time point set. (See also Morris 1987.)
Stockhausen's arguments managed to resolve temporarily one of the many
contradictions and inconsistencies of serial theory. At the same time, they left
unresolved a host of major problems involving perception, notation, traditional
instrumental timbre, and higher-level organization. To untangle and examine
these issues in detail would require another book. Even if these problems could
somehow be magically resolved, it would not automatically ``validate'' compositions made with these techniques, for music will always be more than a game
of logic.
Today it is easier than ever before to compose on all time scales. Yet we
must continue to respect the dierences between them. In his essay on aesthetic
questions in electronic music, the musicologist Carl Dahhaus (1970) criticized
the use of identical methodologies for macro and micro composition. As he
wisely pointed out, serial methods that are already barely decipherable on the
level of notes and phrases disappear into invisibility when applied on the microlevel of tone construction. (See the discussion in chapter 8.)
Henry Cowell's (1930) ideas concerning the relationship between musical
time scales precede those of Stockhausen by almost thirty years. Cowell pointed
out that rhythms, when sped up, become tones. He introduced the concept of
undertones at fractional intervals beneath a fundamental tone, leading to the
notion of a rhythmic undertone. In order to represent divisions of time not

79

The History of Microsound

easily handled by traditional music notation, Cowell proposed a shape note

scheme. He observed that a series of partial frequencies, as seen in spectra,
could be used to build a scale of meters or a scale of tempi, with dierent tempi
running simultaneously and various rates of accelerando and ritardando notated
graphically. He proposed a scale of dynamic stress:
In spite of its importance, there is no adequate notation for dynamics, and the ne distinctions are left to the performer, although such distinctions might well be considered an
essential part of composition. . . . Science can measure the loudness of sound by a number of
well-known means. Since we have a familiar instrument for measuring so delicate a thing as
rate of speednamely, a metronomeit would seem that we should also have some simple
instrument for the measurement of stress. Then we could devise scales of degrees of stress.
This could be done on an additional sta, if desired. (Cowell 1930, pp. 823)

Cowell's writings prove that a comprehensive multiscale view of time can be

formulated quite independently of serial theory.
``. . . . . How time passes . . . . .'' exemplies the disadvantages of discussing
microtemporal composition within the limits of traditional instrumental writing. It is hard to imagine how certain acoustic instruments could be made to
cross between the infrasonic and audio frequencies, where, for example, rhythm
turns into pitch and timbre. Stockhausen does not address the question of
timbre until the end of the article, when he muses on the possibility of an
``ideal instrument.''
As noted earlier the technology of synthesis has greatly advanced since
Stockhausen wrote in 1957. Another technology that has changed dramatically
is that of sound editing and spectrum analysis. Today one can easily probe and
alter the inner microrhythm of musical timbre. Stockhausen's spectral view of
rhythm can be measured now by applying the short-time Fourier transform to
signals in the range @0.06 Hz to 30 Hz, in order to obtain frequency analyses of
rhythm (gure 2.5). Pulsar synthesis, presented in chapter 4, blurs the boundary
between the infrasonic and the audio frequency domains, since the fundamental
frequency envelope can cross between them to create either rhythm or continuous tone.
After Kontakte, Stockhausen discarded impulse generation and processing
techniques. He realized only three more works in the electronic music studio.
Both Telemusik (1966), composed at the studios of the NHK in Tokyo, and
Hymnen (1967) composed in Cologne, reected the composer's interest in
intermodulation based on recordings of world music. Then, after a hiatus of
twenty-three years, Oktophonie (1990) realized in collaboration with his son,

80

Chapter 2

Figure 2.5 Spectrum analysis of infrasonic (rhythmic) frequencies. (a) Fundamental

frequency envelope of a pulsar train. (b) Time-domain view of pulsar train. (c) Infraspectrum of pulsar train.

81

The History of Microsound

focused on the movement in space of long sustained tones produced by commercial synthesizers. None of these works, however, continues or extends the
theoretical principles underlying Kontakte.
Other Assessments of Stockhausen's Temporal Theory
Christopher Koenigsberg (1991) presented a balanced analysis of Stockhausen's
temporal theory and of its critics. His review prompted me to revisit these
contemporaneous critiques. The most heated seem to have been provoked
by Stockhausen's nonstandard terminology. The American acoustician John
Backus wrote:
In physics, a quantum is an undivisible unit; there are no quanta in acoustical phenomena. (Backus 1962, p. 18; see also Backus 1969, p. 280)

Backus had probably not read Gabor's papers. And while certain of his
criticisms concerning terminology are valid, they tend to focus on the details of
a much broader range of musical ideas. The same can be said of Adriaan Fokker's (1962) comments. G. M. Koenig's (1962) response to Fokker's critique
attempts to re-explicate Stockhausen's theories. But Koenig manages to be
even more confusing than the original because of his insistence on defending
its nonstandard terminology. This leads him to some arcane arguments, such as
the convoluted attempt to explain the word ``phase'' (Koenig 1962, pp. 825).
(See also Davies 1964, 1965.)
Milton Babbitt's (1962) proposal on time-point sets, which we have already
mentioned, was published in the same issue of Perspectives of New Music as
the Backus attack on Stockhausen's article. This might have been intentional,
since Babbitt's theory was an alternative proposal for a set-theory approach to
rhythmic structure. The theory of time-point sets, however, left open the question of which durations to use, and was not concerned with creating a unique
scale of equal-tempered durations. Babbitt broached the possibility of microrhythms, but never followed up on this concept.

Microsound in the Analog Domain

The wave model of sound informed the design of early analog electronic
instruments (table 2.1). The Thereminovox, for example, operated in continu-

82

Chapter 2

ous wave emissionalways oscillatingwith one hand controlling the amplitude of the emitted tone. In the hands of a virtuoso, such as Clara Rockmore or
Lydia Kavina, these instruments produce expressive tones, although the duration and density of these tones never approaches the microsonic threshold.
Further examples of a timeless wave model include the sine and pulse generators of the pioneering electronic music studios. These devices were designed
for precise, repeatable, and unchanging output, not for real time performance.
A typical generator might have a vernier dial for frequency, but with the
sweepable frequency range broken into steps. This meant that a continuous
sweep, say from 1 Hz to 10 kHz, was not possible. The amplitude and waveform controls typically oered several switchable settings.
Analog magnetic tape oered a breakthrough for microsound processing. In
discussing the musique concrete of the early 1950s, M. Chion wrote:
Soon the tape recorder, which was used in the musique concrete, would replace the turntable. It allowed editing, which was dicult with the vinyl disc. The possibility of assembling
tight mosaics of sound fragments with magnetic tape denitively launched electroacoustic
music. The rst pieces for tape were ``micro-edited,'' using as their basis sounds that were
reduced to the dust of temporal atoms (Pierre Henry's Vocalises, Boulez's Etudes,
Stockhausen's Etude Concrete). In this ``analytic'' period, one sought the atomic ssion of
sound, and magnetic tape (running at the time at 76 cm/s) was seen as having a tangible
duration that could be cut up ad innitum, up to one hundred parts per second, allowing the
realization of abstract rhythmic speculations that human performers could never play, as in
the Timbres-Durees (1953) of Olivier Messaien. (Chion 1982)

Composers could record the emissions from laboratory generators on tape,

using a potentiometer in the recording chain to vary the amplitude manually.
The duration of the tones could be rearranged by tape cutting and splicing, and
the spectrum could be altered by ltering. Even so, the precision of these operations was, as G. M. Koenig explained, limited:
If the frequency of tuning is 440 cycles per second . . . The individual vibration period thus
lasts 1/440th of a second. But the studio has not a device at its disposal which makes it
possible to open a generator for this length of time, should one want to use a single period.
Even if such a device were available, the tape would still have to be cut o 0.068 of an
inch. . . . It has become possible to assemble timbres from components in the studio, but on
the other hand it is impossible determine their absolute duration at will, because of the
limitations of the ``instrument,'' namely the apparatus in the electronic studio. (Koenig
1959)

Presaging the advent of digital particle synthesis, he goes on:

83

The History of Microsound

If one could get around this obstacle, composition of timbre could be transposed to a time
region in which individual elements would hardly be audible any longer. The sound would
last not seconds . . . but only milliseconds. . . . Instead of ve sounds, we should compose
fty, so that the number of points in the timetable would radically increase. But these
points would not be lled out with sinus tones perceptible as such, but single periods, which
would only be audible en masse, as a uctuating timbre. (Koenig 1959)

Only by the mid 1970s, through the introduction of digital technology, was it
feasible to experiment with microsound in the manner predicted by Koenig.
(See chapter 3.) Digital editing on any time scale did not become possible until
the late 1980s.
The novel sound of analog circuitry was used brilliantly in early works by
Stockhausen, Koenig, and Xenakis to create a new musical world based on
microsonic uctuations. The acoustical signature of analog generators and lters remains a useful resource for twenty-rst century composers. At the same
time, one must recognize the constraints imposed by analog techniques, which
can be traced to a nite class of waveforms and the diculty of controlling their
evolution on a micro level. New analog synthesizers have been introduced into
the marketplace, but they are no more than variations on a well-established
theme. There is little room for further evolution in analog technology.

Microsound in the Instrumental Domain

In his early orchestral pieces Metastasis, Pithoprakta, and Achorripsis, Iannis
Xenakis creates sparse or dense clouds made of brief, irregular notes, especially
string pizzicati. Similar techniques abound in certain compositions of Ligeti
and Penderecki, among others, from the 1960s and 1970s. During this period
the ``stochastic cloud'' was sometimes reduced to a special eect. Although this
was taken up by many other composers, little new has been added.
Detailed micro control of intricate multipart acoustic music is not practical.
Technical, social, and economic obstacles have to be overcome in order to
pursue this path, and those who try to coax unwilling institutions, performers,
and instruments to realize microtemporal music go against the grain of the
music establishment.
A composer whose instrument is the computer does not face these limitations. Thus a more practical strategy for the composer who seeks a link to
traditional sonorities seems to be a mixed approach, combining instrumental

84

Chapter 2

performance with electronic processing (granulation, for example, see chapter

5) and synthetic sound.

Summary
The notion that apparently continuous phenomena can be subdivided into
particles can be traced to the atomistic philosophers of Greek antiquity.
Debates between proponents of the ``wave'' and ``particle'' views in optics and
acoustics have occupied scientists for centuries. These debates were central to
the formation of early modern science.
The contemporary scientic view of microsound dates back to Dennis
Gabor, who applied the concept of an acoustic quantum (already introduced by
Einstein) to the threshold of human hearing. With Meyer-Eppler as intermediary, the pioneering composers Xenakis, Stockhausen, and Koenig injected this
radical notion into music. Xenakis's theory of granular synthesis has proven to
be an especially inspiring paradigm. It has directly inuenced me and many
other composers who have employed granular techniques in their works. Over
decades, a microsonic perspective has gradually emerged from the margins of
musical thought to take its present place as a valuable fountain of compositional ideas.

Granular Synthesis

Theory of Granular Synthesis

Anatomy of a Grain
The Grain Generator
Global Organization of the Grains
Matrices and Screens on the Time-Frequency Plane
Pitch-Synchronous Granular Synthesis
Synchronous and Quasi-Synchronous Granular Synthesis
Pitch and Noise Perception in Synchronous Granular Synthesis
Asynchronous Granular Synthesis
Physical and Algorithmic Models
Streams and Clouds of Granulated Samples
Spectra of Granular Streams
Parameters of Granular Synthesis
Grain Envelope Shape Eects
Experiments in Time Reversal
Grain Duration Eects
Grain Waveform Eects
Frequency Band Eects
Density and Fill Factor
Granular Spatial Eects
Granular Clouds as Sound Objects
Cloud Mixtures

86

Chapter 3

Implementations of Granular Synthesis

The Author's Implementations of Granular Synthesis
Other Implementations of Granular Synthesis
Summary
To stubbornly conditioned ears, anything new in music has always been called noise. But
after all, what is music but organized noises?
Edgard Varese (1962)

Digital sound synthesis techniques inhabit a virtual world more pure and precise than the physical world, and purity and precision have an undeniable
charm in music. In the right hands, an unadorned sine wave can be a lush and
evocative sonority. A measured pulsation can invite emotional catharsis. Synthesis, however, should be able to render expressive turbulence, intermittency,
and singularity; the overuse of precision and purity can lead to sterile music.
Sonic grains, and techniques used to scatter the grains in evocative patterns, can
achieve these results.
This chapter is devoted entirely to granular synthesis (GS). I present its
theory, the history of its implementations, a report on experiments, and an
assessment of its strengths and weaknesses. A thorough understanding of
the principles of granular synthesis is fundamental to understanding the other
techniques presented in this book. This chapter focuses on synthesis with
synthetic waveforms. Since granulation transforms an existing sound, I present the granulation of sampled sound in chapter 5 with other particle-based
transformations.

Theory of Granular Synthesis

The seeds of granular synthesis can be traced back to antiquity, although it was
only after the papers of Gabor and Xenakis that these seeds began to take root
(see chapter 2). A grain of sound is a brief microacoustic event, with a duration
near the threshold of human auditory perception, typically between one thousandth of a second and one tenth of a second (from 1 to 100 ms). Each grain
contains a waveform shaped by an amplitude envelope (gure 3.1).

87

Granular Synthesis

Figure 3.1 Portrait of a grain in the time domain. The duration of the grain is typically
between 1 and 100 ms.

A single grain serves as a building block for sound objects. By combining

thousands of grains over time, we can create animated sonic atmospheres. The
grain is an apt representation of musical sound because it captures two perceptual dimensions: time-domain information (starting time, duration, envelope
shape) and frequency-domain information (the pitch of the waveform within
the grain and the spectrum of the grain). This stands in opposition to samplebased representations that do not capture frequency-domain information, and
abstract Fourier methods, which account only for the frequency domain.
Granular synthesis requires a massive amount of control data. If n is the
number of parameters per grain, and d is the density of grains per second, it
takes n times d parameter values to specify one second of sound. Since n is
usually greater than ten and d can exceed one thousand, it is clear that a global
unit of organization is necessary for practical work. That is, the composer species the sound in global terms, while the granular synthesis algorithm lls
in the details. This greatly reduces the amount of data that the composer must
supply, and certain forms of granular synthesis can be played in real time with
simple MIDI controllers. The major dierences between the various granular
techniques are found in these global organizations and algorithms.

88

Chapter 3

Anatomy of a Grain
A grain of sound lasts a short time, approaching the minimum perceivable
event time for duration, frequency, and amplitude discrimination (Whiteld
1978; Meyer-Eppler 1959; Winckel 1967). Individual grains with a duration less
than about 2 ms (corresponding to fundamental frequencies > 500 Hz) sound
like clicks. However one can still change the waveform and frequency of grains
and so vary the tone color of the click. When hundreds of short-duration grains
ll a cloud texture, minor variations in grain duration cause strong eects in
the spectrum of the cloud mass. Hence even very short grains can be useful
musically.
Short grains withhold the impression of pitch. At 5 ms it is vague, becoming
clearer by 25 ms. The longer the grain, the more surely the ear can hear its
pitch.
An amplitude envelope shapes each grain. In Gabor's original conception,
the envelope is a bell-shaped curve generated by the Gaussian method (gure
3.2a).
1
2
px p ex =2 dx
2p
A variation on the pure Gaussian curve is a quasi-Gaussian envelope (Roads
1978a, 1985), also known as a cosine taper or Tukey window (Harris 1978). This
envelope can be imagined as a cosine lobe convolved with a rectangle (gure
3.2b). It transitions smoothly at the extrema of the envelope while maximizing the eective amplitude. This quality persuaded me to use it in my earliest
experiments with granular synthesis.
In the early days of real-time granular synthesis, it was necessary to use simple line-segment envelopes to save memory space and computation time (Truax
1987, 1988). Gabor (1946) also suggested line-segment envelopes for practical
reasons (gure 3.2c and d). Keller and Rolfe (1998) have analyzed the spectral
artefacts introduced by a line-segment trapezoidal window. Specically, the
frequency response is similar to that of a Gaussian window, with the addition of
comb-shaped spectral eects. Null points in the spectrum are proportional to
the position of the corners of the window.
Figure 3.2e portrays another type of envelope, the band-limited pulse or sinc
function. The sidelobes (ripples) of this envelope impose a strong modulation eect. The percussive, exponentially decaying envelope or expodec grain

89

Granular Synthesis

Figure 3.2 Grain envelopes. (a) Gaussian. (b) Quasi-Gaussian. (c) Three-stage line
segment. (d) Triangular. (e) Sinc function. (f ) Expodec. (g) Rexpodec.

90

Chapter 3

(gure 3.2f ) has proven to be eective in transformations such as convolution

(described in chapter 5). Figure 3.2g depicts the reversed expodec or rexpodec grain. Later we study the strong eect the grain envelope imposes on the
spectrum.
The grain envelope and duration can vary in a frequency-dependent manner
(shorter envelopes for high frequency sounds); such a correlation is characteristic of the wavelet transform (see chapter 6), and of the grainlet synthesis technique described in chapter 4.
The waveform within the grain is an important grain parameter. It can vary
from grain to grain, be a xed waveform that does not change over the grain's
duration, or it can be a time-varying waveform. Typical xed waveforms include the sine wave and sums of sine waves with increasing harmonic content
up to a bandlimited pulse. A time-varying waveform can be generated by frequency modulation or another mathematical technique (Jones and Parks 1988).
The grain waveform can also be a single period extracted from a recorded
sound. This diers from granulation (chapter 5), which scans across a long
sampled waveform, extracting many dierent grains over time.
Other parameters of the grain include its duration, the frequency of its
waveform, its amplitude coecient, and its spatial location. We examine these
parameters later.
The Grain Generator
In its most basic form, the grain generator is a simple digital synthesis instrument. Its circuit consists of a wavetable oscillator whose amplitude is controlled
by a Gaussian envelope. The output of the oscillator goes to a spatial panner
(gure 3.3).
As a general principle of synthesis, we can trade o instrument complexity
for score complexity. A simple instrument suces, because the complexity of
the sound derives from the changing combinations of many grains. Hence
we must furnish a massive stream of control data for each parameter of the
instrument.
Despite the simplicity of the instrument, we gain two spectral controls for
free, owing to the specic properties of the micro time scale. Specically,
variations in the grain duration and the grain envelope aect the spectrum of
the resulting signal (as a later of this book section claries).
Of course, we can always make the granular synthesis instrument more complex, for instance by adding a local frequency control, per-grain reverberation,

91

Granular Synthesis

Figure 3.3 The simplest grain generator, featuring a Gaussian grain envelope and a
sinusoidal grain waveform. The grains can be scattered to a position in N channels of
output.

multichannel output, etc. Chapter 4 describes several extensions to the basic

granular instrument.
Global Organization of the Grains
The main forms of granular synthesis can be divided into six types, according to
how each organizes the grains. They are:
1 Matrices and screens on the time-frequency plane
1 Pitch-synchronous overlapping streams
1 Synchronous and quasi-synchronous streams
1 Asynchronous clouds
1 Physical or abstract models
1 Granulation of sampled sound

92

Chapter 3

Matrices and Screens on the Time-Frequency Plane

The Gabor matrix, shown in chapter 2, is the original time-frequency matrix for
sound, based on the analysis of an existing sound. In the same general family
are the analyses produced by the short-time Fourier transform (STFT) and the
wavelet transform (WT), presented in chapter 6. These operations transform a
time-domain signal into a frequency-domain representation that is quantized in
both the time and frequency dimensions, creating a two-dimensional matrix.
Frequency-domain matrices oer many opportunities for sound transformation.
A related organizational scheme is Xenakis's (1960, 1971) notion of screens
(also described in chapter 2). Each screen is like a snapshot of a microsound. It
represents a Gabor matrix at a specic moment, divided into cells of amplitude
and frequency. Like the frames of a lm, a synchronous sequence of screens
constitutes the evolution of a complex sound. Rather than starting from an
analyzed sound, Xenakis proposed to ll the screen with grains by means of
stochastic algorithms. Another proposal suggested that the grains be generated
from the interaction of cellular automata (Bowcott 1989; Miranda 1998).
Pitch-Synchronous Granular Synthesis
Pitch-synchronous granular synthesis (PSGS) is an ecient analysis-synthesis
technique designed for the generation of pitched sounds with one or more
formant regions in their spectra (De Poli and Piccialli 1991). It begins with a
spectrum analysis of a sound. Each time-frequency cell corresponds to a grain.
As a preparation for resynthesis, at each grain boundary along the frequency
axis, a standard algorithm derives the coecients for a lter. The impulse response of this lter corresponds to the frequency response of the cell. At each
grain boundary along the time axis, a pitch detection algorithm determines the
fundamental pitch period. In resynthesis, a pulsetrain at the detected frequency
drives a bank of parallel minimum-phase nite impulse response lters. The
musical signal results from the excitation of the pulsetrain on the weighted sum
of the impulse responses of all the lters. At each grain time frame, the system
emits a waveform that is overlapped with the previous grain to create a
smoothly varying signal. An implementation of PSGS by De Poli and Piccialli
features several transformations that can create variations of the original
sound. A focus of their implementation was the use of data reduction techniques to save computation and memory space. See De Poli and Piccialli (1991),
and Cavaliere and Piccialli (1997) for details.

93

Granular Synthesis

Synchronous and Quasi-Synchronous Granular Synthesis

Granular streams appear naturally from iterative sound productionany kind of roll or
trill on drums, percussion, or any sounding material. They are produced vocally by rolled
``r'' sounds. (Wishart 1994)

In synchronous granular synthesis (SGS), sound results from one or more

streams of grains. Within each stream, one grain follows another, with a delay
period between the grains. ``Synchronous'' means that the grains follow each
other at regular intervals.
An excellent use of SGS is to generate metric rhythms, particularly when the
grain emissions are sparse per unit of time. The density parameter controls the
frequency of grain emission, so grains per second can be interpreted as a frequency value in Hertz. For example, a density of 1 grain per second indicates
that a grain is produced every second. Synchronous densities in the range of
about 0.1 to 20 grains per second will generate metrical rhythms. When densities change over time, the listener hears accelerandi and rallentandi.
At higher densities, the grains fuse into continuous tones. Here is found the
sweeter side of granular synthesis. These tones have a strong fundamental frequency, and depending on the grain envelope and duration, may also exhibit
sidebands. The sidebands may sound like separate pitches or they may blend
into a formant peak. At certain settings, these tones exhibit a marked vocal-like
quality. In these cases, SGS resembles other techniques such as FOF and Vosim
synthesis. Chapter 4 describes these particle-based formant synthesis methods.
The formant shape and strength depend greatly on the grain duration and
density which also, under certain conditions, aect the perceived fundamental
frequency. We explore this in more detail in the next section.
In quasi-synchronous granular synthesis (QSGS), the grains follow each other
at unequal intervals, where a random deviation factor determines the irregularity. If the irregularity is great, the sounds produced by this method become
similar to those produced by asynchronous granular synthesis. SGS and QSGS
are well-adapted to real-time implementations.
Pitch and Noise Perception in Synchronous Granular Synthesis
The fragility of the illusion of pitch is made apparent in SGS. The perceived
pitch of a granular stream depends primarily on interactions among three
periodicities:

94

Chapter 3

a is the period corresponding to the frequency of the waveform in the grain

b is the period corresponding to the grain envelope
c is the period corresponding to the density, or rate of synchronous grain
emission
One or another of these factors may override the others in determining the
perceived pitch. In certain cases, modulations caused by their interaction may
render the pitchand especially the specic octaveambiguous.
Figure 3.4 illustrates these eects. Consider a tone made up of a series of
10 ms grains, where each grain contains two periods of a 200 Hz sine wave.
Assume, as in gure 3.4a, that the density of the grains is 50 per second. Here
we have:
a 5 ms
b 10 ms
c 20 ms
As gure 3.4a shows, the 10 ms gap between b and c means that there is a
dead interval between successive grains, leading to a modulation eect with its
associated sidebands. The perceived pitch is a buzzy 100 Hz.
A linear increase in grain density (from 50 to 100 grains per second in the
above case) causes a pitch doubling eect. The perceived pitch is now 200 Hz.
Here the three variables take the values:
a 5 ms
b 10 ms
c 10 ms
In gure 3.4c, the grain density is 200 grains per second. The variables take
these values:
a 5 ms
b 10 ms
c 5 ms
Now we have a pure sinusoidal tone. Only one period of the waveform can
unfold within the grain repetition period c, and the inuence of the grain envelope b is diminished.

95

Granular Synthesis

Figure 3.4 Inuence of grain density on pitch. The waveforms in (a) through (e) last
59 ms. (a) 50 grains/sec. (b) 100 grains/sec. (c) 200 grains/sec. (d) 400 grains/sec.
(e) 500 grains/sec. (f ) Plot of a granular stream sweeping from the infrasonic frequency
of 10 grains/sec to the audio frequency of 500 grains/sec over thirty seconds.

When the grain density increases to 400 grains per second (gure 3.4d), the
perceived pitch doubles to 400 Hz. This is due to the increasing frequency of
wavefronts (as in the well-known Doppler shift eect). Notice that the amplitude of tone diminishes after beginning, however, because the density period c
is less than the waveform period a. Only the rst few samples of the product of
the sine wavetable and the grain envelope are being repeated, resulting in a lowamplitude signal.
Finally, at a density of 500 grains per second (gure 3.4e), the signal has
almost no amplitude. It is reading only the rst few samples of the sinusoid,
which are near zero.

96

Chapter 3

Figure 3.4f shows the amplitude prole of a granular stream that sweeps
from 10 grains per second to 500 grains per second over thirty seconds. Notice
the diminution of amplitude due to the eect shown in gure 3.4e.
Besides pitch changes, other anomalies, such as phase cancellation, can occur
when the grain density and envelope duration are at odds with the frequency of
the grain waveform.
Even the impression of synchronicity can be undermined. If we widen the
frequency limits of a dense synchronous stream slightly, the result quickly truns
into a noiseband. The fact that the grain emissions are regular and the frequency changes at regular intervals (for example, every 1 ms), does not alter
the general impression of noise. The eect is similar to that produced by asynchronous granular synthesis, described next.
Asynchronous Granular Synthesis
Asynchronous granular synthesis (AGS) abandons the concept of linear streams
of grains. Instead, it scatters the grains over a specied duration within regions
inscribed on the time-frequency plane. These regions are cloudsthe units with
which the composer works. The scattering of the grains is irregular in time,
being controlled by a stochastic or chaotic algorithm. The composer may specify
a cloud with the following parameters:
1. Start-time and duration of the cloud
2. Grain durationmay vary over the duration of the cloud
3. Density of grains per second, with a maximum density depending upon the
implementation; density can vary over the duration of the cloud
4. Frequency band of the cloud; specied by two curves forming high and low
frequency boundaries within which grains are scattered; alternatively, the
frequency of the grains in a cloud can be restricted to a specic set of pitches
5. Amplitude envelope of the cloud
6. Waveform(s) within the grains
7. Spatial dispersion of the cloud, where the number of output channels is implementation-specic
The grain duration (2) can be a constant (in milliseconds), or a variable that
changes over the course of a cloud. (It can also be correlated to other parameters, as in the grainlet synthesis described in chapter 4.) Grain duration can also

97

Granular Synthesis

be derived as a random value between an upper and a lower boundary set by

the user. The next section explains the eects of dierent grain durations in
more detail.
Parameter (3), grain density, species the number of grains per unit of time.
For example, if the grain density is low, then only a few grains are scattered at
random points within the cloud. If the grain density is high, grains overlap to
create rich, complex spectra.
Parameter (6) is one of the most exible cloud parameters, since each grain
may have a its own waveform.
Physical and Algorithmic Models
Physical modeling (PhM) synthesis starts from a mathematical description of
acoustic sound production (Roads 1996; Fletcher and Rossing 1991). That is,
the equations of PhM describe the mechanical and acoustical behavior of an
instrument as it is played. An example of physical modeling applied to granular synthesis is Perry Cook's Physically Informed Stochastic Event Modeling
(PhISEM). This suite of programs simulates the sounds of shaken and scraped
percussion such as maracas, sekere, cabasa, bamboo windchime, tambourine,
sleighbells, and guiro (Cook 1996, 1997). See more about this technique in chapter 4. Going beyond traditional instruments, Keller and Truax created sound
models of such physical processes as the bounce of a metallic ball and the rush
of a stream of water (Keller and Truax 1998). Physical models of granular
processes could be taken still further. A large body of scientic literature centers on granular processes such as grain mixing, granular ow, grain vibration
patterns, and grain and uid interactions. This literature appears in research
periodicals such as Granular Matter, Powders and Grains, Powder Technology,
Journal of Fluid Mechanics, Physical Review, Journal of Applied Mechanics, as
well as such Internet web sites as www.granular.com.
Going beyond emulations of the natural world, one can also develop models
of virtual worlds through abstract algorithms. To cite an example, Alberto de
Campo (1998) proposed a method of grain scattering in time based on a recursive substitution algorithm. Another idea would be using chaotic functions
to scatter the grains in time (Gleick 1988; Holden 1986; Moon 1987). Chaotic
functions produce dierent results from scattering algorithms based on pseudorandom algorithms. The values produced by pseudorandom number generators tend to be uniform and adirectional, tending toward the mean. To make
them directional, they must be ltered through stochastic weightings. In con-

98

Chapter 3

trast, chaotic functions vacillate between stable and unstable states, between
intermittent transients and full turbulence (Di Scipio 1990, 1997b; Gogins 1991,
1995; Miranda 1998). The challenge is to set up a musically compelling mapping between chaotic behavior and the synthesis parameters.
Streams and Clouds of Granulated Samples
The granulation of sampled sounds is a powerful means of sound transformation. To granulate means to segment (or window) a sound signal into grains, to
possibly modify them in some way, and then to reassemble the grains in a new
time order and microrhythm. This might take the form of a continuous stream
or of a statistical cloud of sampled grains.
The exact manner in which granulation occurs will vary from implementation to implementation. Chapter 5 includes a major section on granulation so
here we shall limit the discussion to noting, that granulation can be controlled
by any of the global control structures described above.
Spectra of Granular Streams
When the intervals between successive grains are equal, the overall envelope of
a stream of grains forms a periodic function. Since the envelope is periodic, the
signal generated by SGS can be analyzed as a case of amplitude modulation or
AM. AM occurs when the shape of one signal (the modulator) determines the
amplitude of another signal (the carrier). From a signal processing standpoint,
we observe that for each sinusoidal component in the carrier, the periodic envelope function contributes a series of sidebands to the nal spectrum. (Sidebands are additional frequency components above and below the frequency of
the carrier.) The sidebands separate from the carrier by a distance corresponding to the inverse of the period of the envelope function. For grains lasting
20 ms, therefore, the sidebands in the output spectrum will be spaced at 50 Hz
intervals. The shape of the grain envelope determines the precise number and
amplitude weighting of these sidebands.
The result of modulation by a periodic envelope is that of a formant surrounding the carrier frequency. That is, instead of a single line in the spectrum
(a single frequency), the spectrum looks like a sloping peak (a group of frequencies around the carrier). In the case of a bell-shaped Guassian envelope,
the spectrum is similarly bell-shaped. In other words, for a Gaussian envelope,
the spectrum is an eigenfunction of the time envelope.

99

Granular Synthesis

When the delay interval between the grains is irregular, perfect grain synchronization disappears. The randomization of the onset time of each grain
leads to a controllable thickening of the sound spectruma ``blurring'' of the
formant structure (Truax 1988).
In its simplest form, the variable-delay method is similar to amplitude modulation using low-frequency colored noise as a modulator. In itself, this is not
particularly new or interesting. The granular representation, however, lets us
move far beyond simple noise-modulated AM. We can simultaneously vary
several other parameters on a grain-by-grain basis, such as grain waveform,
amplitude, duration, and spatial location. On a global level, we can also dynamically vary the density of grains per second, creating a variety of scintillation eects.
Parameters of Granular Synthesis
Research into sound synthesis is governed by aesthetic goals as much as by
scientic curiosity. Some of the most interesting synthesis techniques have
resulted from applied practice, rather than from formal theory. Sound design
requires taste and skill and at the experimentation stage, musical intuition is the
primary guide.
Grain Envelope Shape Eects
Of Loose Atomes
In every Braine loose Atomes there do lye,
Those which are Sharpe, from them do Fancies ye.
Those that are long, and Aiery, nimble be.
But Atomes Round, and Square, are dull, and sleepie.
(Margaret Cavendish 1653)

This section presents empirical reports on the eects caused by manipulating

the grain envelope, duration, waveform, frequency, band, density, and spatial
parameters.
Referring back to gure 3.2, the classical grain envelope is the bell-shaped
Gaussian curve (gure 3.2a). This is the smoothest envelope from a mathematical point of view. A quasi-Gaussian (Tukey) envelope retains the smooth
attack and decay but has a longer sustain portion in the envelope and so
increases its perceived amplitude (gure 3.2b). Compared to a pure Gaussian of
the same duration, the quasi-Gaussian broadens the spectrum. Its highest side-

100

Chapter 3

lobe is only 18 dB down, as opposed to 42 dB for a pure Gaussian curve

(Harris 1978). The band-limited pulse or sinc function imposes a strong modulation eect (gure 3.2e). I have used it myself to create ``bubbling'' or ``frying''
clouds.
I have carried out numerous experiments using grain envelopes with a sharp
attack (typically less than 10 ms) and an exponential decay. These are the
expodec grains (gure 3.2f ). The percussive attack of the expodec articulates
the rhythmic structure. As chapter 5 describes, clouds of expodec grains can be
especially useful as impulse responses for convolution with other sounds. Reversed expodec or rexpodec grains feature a long attack envelope with a sudden
decay (gure 3.2g). Granulated concrete sounds appear to be ``reversed'' when
they are played with rexpodec grains, even though they are not.
While keeping the envelope shape constant, a change in the grain duration
has a strong eect on the spectrum. Furthermore, the grain envelope and duration can vary in a frequency-dependent manner (shorter envelopes for high
frequency sounds); such a correlation is characteristic of the wavelet transform
and the grainlets.
Experiments in Time Reversal
The concept of time has been thoroughly twisted by modern theoretical physics
(Kaku 1995). Barry Truax (1990b) drew an analogy between the world of particle physicsin which time appears to be reversible at the quantum leveland
granular synthesis. According to his hypothesis, if a grain is reversed in time,
it should sound the same. Moreover, granular synthesis textures should also
be reversible. Such a position would also follow on from Trevor Wishart's
assertion:
Although the internal structure of sounds is the cause of what we hear, we do not resolve
this internal structure in our perception. The experience of a grain is indivisible. (Wishart
1994)

Under special circumstances, all of this is quite true. But if we loosen any one
of a number of constraints, time reversibility does not hold. For it to hold at the
micro scale, the grain envelope must be symmetrical. This, then, excludes
asymmetric techniques such as FOF grains (Rodet 1980), trainlets (chapter 4),
expodec, or rexpodec grains. The grain waveform must not alter in time, so
excluding techniques such as the time-varying FM grains (Jones and Parks
1988), long glissons (chapter 4), or grains whose waveform derives from a time-

101

Granular Synthesis

varying sampled sound. With texture a grain stream or cloud is time-reversible

only if it is stationary in the statistical sense, meaning that its overall amplitude
and spectral envelopes are symmetrical, its internal density is constant, and the
waveform of all the grains is similar.
Grain Duration Eects
The duration of the grains inside a cloud has profound eects on the resulting
audio signal. Within clouds, there are four classes of grain durations:
1. Constant duration

the duration of every grain in the cloud is the same.

2. Time-varying duration

the grain duration varies as a function of time.

3. Random duration the duration of a grain is random between an upper and

lower duration boundaries.
4. Parameter-dependent duration the duration of a grain is tied to its fundamental frequency period, as it is in synthesis with wavelets, or any other
parameter, as in the grainlet synthesis.
Regarding constant durations, early estimates of the optimum grain duration
varied from 10 ms (Gabor 1946, 1947) to 60 ms (Moles 1968).
The grain envelope contributes an amplitude modulation (AM) eect. The
modulation spawns sidebands around the carrier frequency of the grain at
intervals of the envelope period. If the grain duration is D, the center frequency
of the AM is 1=D. In an asynchronous cloud, the AM sounds like an aperiodic
uttering tremolo when D is around 100 ms (table 3.1).
Table 3.1 Eects of grain durations in asynchronous granular synthesis
Grain duration

Frequency of
modulation

200 msec
500 msec
1 ms
10 ms
50 ms
100 ms
200 ms

5 KHz
2 KHz
1 KHz
100 Hz
20 Hz
10 Hz
5 Hz

Perceived eect
Noisy particulate disintegration
Loss of pitch
Fluttering, gurgling
Stable pitch formation
Aperiodic tremolo, jittering

102

Chapter 3

Figure 3.5 Comparison of grain spectra produced by a 7 ms grain duration (top) versus
a 29 ms grain duration (bottom). Notice the narrowing of the spectrum as the duration
lengthens.

The laws of micro-acoustics tell us that the shorter the duration of a signal,
the greater its bandwidth. Thus the width of the frequency band B caused by
the sidebands is inversely proportional to the duration of the grain D (gure
3.5).
A dramatic eect occurs when the grain duration is lowered to below the
period of the grain waveform. This results in a signal that is entirely unipolar in
energy, which is a byproduct of the ratio of the grain duration to the fundamental frequency period Pf of the grain waveform, or D=Pf . The eect is
caused by an incomplete scan of the wavetable, where the waveform starts in
either the positive or the negative quadrant. It occurs whenever D=Pf is less
than 1.0. In the specic case of a 1 ms grain with a fundamental frequency of
500 Hz, the ratio is 0:001=0:002 1=2.
To completely represent one period of a given frequency, the grain duration
must be at least equal to the frequency period. If we took this criterion as a
standard, grains could last no less than 50 ms (corresponding to the period of
20 Hz) for low frequency signal energy to be captured completely. As it happens however, much shorter grains can represent low frequency signals, but this
short grain duration introduces modulation products. Our experiments show
that grains shorter than 5 ms tend to generate particulated clouds in which a
sense of center-pitch is still present but is diused by noise as the frequency
descends.

103

Granular Synthesis

Grain Waveform Eects

One of the most interesting features of granular synthesis is that one can insert
any waveform into a grain. The waveform can vary on a grain-by-grain basis.
This makes possibile micro-animated textures that evolve directionally over
time or simply scintillate from the eects of constantly changing grain waveforms.
The simplest grain waveforms are the xed synthetic types: the sine, saw,
square, and sinc (band-limited impulse). In early experiments, I used ten
synthetic waveforms created by adding one to ten sine waves in a harmonic
relationship.
Interspersing dierent waveforms in a single cloud leads to cloud color type
(Roads 1991). Three possibilities for cloud color type are:
1 monochrome containing a single waveform
1 polychrome containing two or more waveforms
1 transchrome the grain waveform evolves from one waveform to another
over the duration of the cloud
For a monochrome cloud, we stipulate a single wavetable for the entire
cloud. For a polychrome cloud, we specify two or more waveforms which can
scatter uniformly in time or according to a time-varying tendency curve. For a
transchrome cloud, if we specify a list of N waveforms, the cloud mutates from
one to the next, through all N over its duration.
So far we have discussed waveform variations on the time scale of clouds.
But even within a single grain, the waveform may be varying in time. The grain
waveform could be generated by time-varying frequency modulation, for
example. Since the duration of the grain is brief, however, such techniques tend
to result in noisy, distorted textures unless the modulating frequencies and the
amount of modulation are strictly controlled.
As a practical aside, it has been necessary to use the standard 44.1 or 48 kHz
sampling rates for software and hardware compatibility in recording, synthesis,
and playback. These sampling rates provide little ``frequency headroom,'' and
one must be aware that when the fundamental frequency of a grain is high and
the waveform is complex, aliasing can occur. To avoid this, one can constrain
the choice of waveform depending on the fundamental frequency, particularly
in the region above half of the Nyquist frequency (11.025 or 12 kHz, depending

104

Chapter 3

on the sampling rate). Above these limits, waveforms other than sine cause
foldover. For this reason, higher sampling rates are better for digital synthesis.
The grain waveform can also be extracted from a sampled sound. In this
case, a single extracted waveform is fed to an oscillator, which reads the waveform repetitively at dierent frequencies. In Cloud Generator, for example, the
extracted waveform constitutes the rst 2048 samples (46 ms) of a selected
sound le (see the appendix). This diers from granulation, which extracts
many dierent segments of a long sample le. See chapter 5.
Frequency Band Eects
Frequency band parameters limit the fundamental frequencies of grain waveforms. Within the upper and lower boundaries of the band, the grain generator
scatters grains. This scattering can be aligned to a frequency scale or to random
frequencies. When the frequency distribution is random and the band is greater
than a small interval, the result is a complex texture, where pitch is ambiguous
or unidentiable. The combined AM eects of the grain envelope and grain
density strongly inuence pitch and spectrum.
To generate harmonic texture, we can constrain the choice of fundamental
frequency to a particular set of pitches within a scale. We distinguish two
classes of frequency specications:
Cumulus The frequencies of the grains scatter uniformly within the upper and
lower bounds of a single band specied by the composer.
Stratus

The frequencies of the grains align to a set of specied frequencies.

Figure 3.6 depicts a variety of frequency band specications for granular

clouds. When the band centers on a single frequency (gure 3.6a), the cloud
produces a single pitch. The frequency can be changed to create a glissando
(gure 3.6b). A stratus cloud contains multiple frequency specications (gure
3.6c). With sampled soundles, one can achieve the harmonic eect of a stratus
cloud by keeping a database of tones at all the pitches from a desired scale, or
by pitch-shifting in conjunction with granulation. When the band is wider than
a single pitch, grains scatter randomly between the upper and lower boundaries
of a cumulus cloud (gure 3.6d). When the initial and nal bands are dierent,
the shape of the cumulus band changes over time (gure 3.6e). In the most
exible case, two time-varying curves shape the bandlimits of the cloud (gure
3.6f ).

105

Granular Synthesis

Figure 3.6 Frequency band specications. (a) The band centers on a single frequency.
(b) The center frequency changes over time, creating a glissando eect. (c) Stratus cloud
with several frequencies. (d) Cumulus cloud where the grains scatter randomly between
the upper and lower boundaries. (e) The shape of the cumulus band changes over time.
(f ) Time-varying curves shape the bandlimits of the cumulus cloud.

Density and Fill Factor

``Density'' is the number of grains per second. If this specication is not linked
with the grain duration, however, it tells us little about the resulting texture.
Grain duration and density combined produce texture.
A one-second cloud containing twenty 100 ms grains is continuous and opaque, whereas a cloud containing twenty 1 ms grains is sparse and transparent.
The dierence between these two cases is their ll factor (FF ). The ll factor of
a cloud is the product of its density and its grain duration in seconds (D). In the
cases just cited, the ll factor of the rst cloud is 20  0:1 2, and of the second
cloud 20  0:01 0:2. These are simple cases, where the density and grain duration are constants, in practice grain density and grain duration can vary over
the duration of the cloud. In this case we derive the average density and the
average ll factor, calculated as the mean between any two extremes. These
measurements provide these descriptors of ll factor:
1 Sparse FF < 0:5, more than half the cloud is silence
1 Covered FF @ 1:0, the cloud is more-or-less lled by sonic grains
1 Packed FF > 1:0, the cloud is lled with overlapping grains

106

Chapter 3

In asynchronous granular synthesis, the starting time of a grain is random.

One cannot guarantee that fty 20 ms grains will completely ll a one-second
cloud. Some grains may overlap, leaving silences at other points in the
cloud. To create what we hear as a solid cloud, a good rule of thumb is to
set the density per second of the cloud to at least 2=D. Hence, for 20 ms
grains, it takes about 100 to cover a one-second cloud. Tiny gaps (less than
about 50 ms) do not sound as silences, but rather as momentary uctuations of
amplitude.
For a typical grain duration of 25 ms, we can make the following observations concerning grain density as it crosses perceptual thresholds:
<15 grains per secRhythmic sequences.
1525 grains per secFluttering, sensation of rhythm disappears. If the cloud
is asynchronous, we hear intermittencies.
2550 grains per secGrain order disappears. Upper and lower frequency
bounds can be inferred. In a synchronous cloud, the perception of synchronicity
evaporates if the upper and lower frequency bounds extend beyond several semitones. As the density increases, we no longer perceive an acceleration of tempo
in the grain emissions, but rather we feel an increase in the ow of grains.
50100 grains per secTexture band. If the bandwidth is greater than a semitone, we cannot discern individual frequencies.
>100 grains per secContinuous sound mass. No space between grains. In
some cases resembles reverberation.
Density and frequency band eects are also synergistic, and, depending on
the grain density, the musical results of the band parameter will dier. For
sparse, pointillist eects, for example, where each grain is heard as a separate
event, keep the grain density to less than 0:5=D, where D is grain duration
in seconds. So, for a grain duration of 20 ms, the density should be less than
25 grains per sec (0.5/0.02).
By increasing the grain density, we enrich the texture, creating eects that
depend on the bandwidth.
1 Narrow bands and high densities generate pitched streams with formant
spectra.
1 Medium bands (e.g., intervals of several semitones) and high densities generate turgid colored noise.

107

Granular Synthesis

1 Wide bands (e.g., an octave or more) and high densities generate massive
clouds of sound.
As we have seen, in the section on grain duration eects, another way to modify
the bandwidth of a cloud is by changing the grain duration parameter.
Granular Spatial Eects
Granular synthesis calls for multichannel output, with an individual spatial
location for each grain. If the cloud is monaural, with every grain in the
same spatial position, it is spatially at. In contrast, when each grain scatters to
a unique location, the cloud manifests a vivid three-dimensional spatial morphology, evident even in a stereophonic conguration.
From a psychoacoustical point of view, the listener's perception of the spatial
position of a grain or series of grains is determined by both the physical properties of the signal and the localization blur introduced by the human auditory
system (Blauert 1997). Localization blur means that a point source sound produces an auditory image that spreads out in space. For Gaussian tonebursts,
the horizontal localization blur is in the range of 0.8 to 3.3 , depending on the
frequency of the signals (Boerger 1965). The localization blur in the median
plane (starting in front, then going up above the head and down behind) is
greater, on the order of 4 for white noise and becoming far greater (i.e. less
accurate) for purer tones. (See Boerger 1965 for a study of the spatial properties
of Gaussian grains.)
Taking localization blur into account, one can specify the spatial distribution
of the grains in one of two ways: as an envelope that pans across N channels, or
as a random dispersion of grains among N channels. Random dispersion is
especially eective in the articulation of long grains at low densities.
Chapter 5 presents more on the spatial eects made possible through particle
scattering and other techniques.
Granular Clouds as Sound Objects
A cloud of grains may come and go within a short time span, for example, less
than 500 ms. In this case, a cloud of grains forms a tiny sound object. The inner
structure of the cloud determines its timbral evolution. I have conducted numerous experiments in which up to fty grains were generated within a time
span of 20 to 500 ms. This is an eective way to construct singular events that
cannot be created by other means.

108

Chapter 3

Cloud Mixtures
A granular composition is a ow of multiple overlapping clouds. To create such
textures, the most exible strategy is rst to generate each individual cloud.
Then to mix the clouds to precisely order and balance their ow in time. To
create a polychrome cloud texture, for example, several monochrome clouds,
each with a dierent grain waveform are superimposed in a mixing program.
It is easy to granulate a sound le and take the results ``as is.'' A more sophisticated strategy is to take the granulation as a starting point. For example,
one can create a compound cloudone with an interesting internal evolutionby carefully mixing several granulated sound les.
Mixing is also eective in creating rhythmic structures. When the density of a
synchronous cloud is below about 20 Hz, it creates a regular metric pulse. To
create a polyrhythmic cloud, one can generate several clouds at dierent densities, amplitudes, and in dierent frequency regions to stratify the layers.

Implementations of Granular Synthesis

This section surveys the history of implementations of granular synthesis on
computers. It begins with my own rst experiments, going on to cover a variety
of implementations since.
The Author's Implementations of Granular Synthesis
My involvement with granular synthesis dates back to May of 1972, when I
participated in Iannis Xenakis's workshop on music and mathematics at
Indiana University. The workshop was based on his book Formalized Music
(Xenakis 1971, 1992). One chapter of this book described a theoretical approach
to sound synthesis based on ``elementary sonic particles:''
A complex sound may be imagined as a multicolored rework in which each point of light
appears and instantaneously disappears against a black sky . . . A line of light would be
created by a suciently large multitude of points appearing and disappearing instantaneously. (Xenakis 1992 pp. 434)

This description intrigued me, but there were no sounds to hear. Granular
synthesis remained a theoretical topic at the workshop. Maestro Xenakis took
us to the campus computing center to show us experiments in stochastic wave-

109

Granular Synthesis

form generation (also described in his book), but he never realized granular
synthesis on a computer.
Later that year, I enrolled as a student in music composition at California
Institute of the Arts. During this period, I also studied mathematics and computer programming with Leonard Cottrell. For the next two years, I wrote
many programs for the Data General Nova 1200, a minicomputer at the Institute. Thus included software for stochastic processes and algorithmic composition based on Xenakis's formulas (Roads 1992a). I spent much time testing the
formulas, which fostered in me a deeper understanding of probability theory.
The Nova 1200 was limited, however. It lacked memory and had no digital
audio converters. Its only peripheral was a teletype with a paper tape punch for
storing and reading programs. Digital sound synthesis was out of the question.
In March 1974, I transferred to the University of California, San Diego
(UCSD), having learned of its computer sound synthesis facilities. Bruce
Leibig, a researcher at UCSD, had recently installed the Music V program
(Mathews 1969) on a mainframe computer housed in the UCSD Computer
Center. The dual-processor Burroughs B6700 was an advanced machine for its
day, with a 48-bit wordlength, virtual memory, digital tape storage, and support for parallel processing. A single language, Extended Algol, provided access
to all levels of the system, from the operating system to the hardware. This is
not to say that music synthesis was easy; because of the state of input and output technology, the process was laborious.
The Burroughs machine could not produce sound directly. It could, however,
write a digital tape that could be converted to sound on another computer, in
this case a Digital Equipment Corporation (DEC) PDP-11/20, housed on
campus at the Center for Music Experiment (CME). Bruce Leibig wrote the
PAL-11 assembly language code that performed the digital-to-analog conversion. This important programming work laid the foundation for my research. I
enrolled in an Algol programming course oered by the computer science
department. There were no courses in computer sound synthesis, but with help
from Bruce Leibig, I learned the Music V language. We programmed on
punched paper cards, as there were no interactive terminals.
Owing to storage limitations, my sound synthesis experiments were limited to
a maximum of one minute of monaural sound at a sampling rate of 20 kHz. It
took several days to produce a minute of sound, because of the large number of
steps involved. The UCSD Computer Center scheduled sound calculations for
the overnight shift. So I would submit a box of punched cards to a computer
operator and return the next day to collect a large digital tape reel containing

110

Chapter 3

the previous evening's data. In order to convert this data into sound, I had rst
to transfer it from the tape to a disk cartridge. This transfer involved setting up
an appointment at the Scripps Institute of Oceanography. Surrounded by the
pungent atmosphere of the squid tanks of the Neurology Computing Laboratory, I transferred the contents of the tape. Then I would take the disk cartridge
to CME and mount it on the DEC minicomputer. This small computer, with a
total of 28 kbytes of magnetic-core RAM, had a single-channel 12-bit digitalto-analog converter (DAC) designed and built by Robert Gross. The digital-toanalog converter truncated the four low-order bits of the 16-bit samples.
After realizing a number of short etudes with Music V, in December 1974 I
tested the rst implemention of asynchronous granular synthesis. For this experiment, called Klang-1, I typed each grain specication (frequency, amplitude, duration) onto a separate punched card. A stack of about eight hundred
punched cards corresponded to the instrument and score for thirty seconds of
granular sound. Following this laborious experience, I wrote a program in
Algol to generate grain specications from compact, high-level descriptions of
clouds. Using this program, I realized an eight-minute study in granular synthesis called Prototype. Chapter 7 describes these studies in detail. (See also
Roads 1975, 1978a, 1985c, 1987.)
In 1980, I was oered a position as a Research Associate at the Experimental
Music Studio at the Massachusetts Institute of Technology. The computing
environment centered on a Digital Equipment Corporation PDP-11/50 minicomputer (16-bit word length) running the UNIX operating system. There I
implemented two forms of granular synthesis in the C programming language.
These programs generated data that could be read by the Music 11 sound synthesis language. The Csound language (Boulanger 2000; Dodge and Jerse 1997;
Vercoe 1993) is a superset of Music 11. The initial tests ran at a 40 kHz sampling rate, and used 1024-word function tables for the waveforms and envelopes. The 1980 implementation generated a textual score or note-list for a
sinusoidal granular synthesis oscillator. The second, 1981, implementation at
MIT granulated sampled sound les using the soundin unit generator of Music
11. I implemented gestures such as percussion rolls by granulating a single
stroke on a snare drum or cymbal. Due to the limitations of the Music 11 language, however, this version was constrained to a maximum density of thirtytwo simultaneous grains.
An important transition in technology took place in the 1980s with the introduction of personal computers. By 1988, inexpensive computers (less than

111

Granular Synthesis

$5000 for a complete system including audio converters) had become powerful
enough to support stereo 16-bit, 44.1 kHz audio synthesis. In 1988, I programmed new implementations of granular synthesis and granulation of sampled soundles for the Apple Macintosh II computer in my home studio (Roads
1992c, d). I called these C programs Synthulate and Granulate, respectively.
For playback, I used the Studer Dyaxis, a digital audio workstation with good
16-bit converters attached to the Macintosh II. My synthesis programs worked
with a version of the Music 4C language, which I modied to handle the large
amounts of data associated with granular synthesis. Music 4C (Gerrard 1989)
was a C-language variant of the venerable Music IVBF language developed in
the 1960s (Mathews and Miller 1965; Howe 1975). I revised the synthesis programs in 1991 while I was at the Kunitachi College of Music in Tokyo. After
moving to Paris in 1992, I modied the grain generator to work with instruments that I wrote for the Csound synthesis language (Boulanger 2000). The
revised programs ran on a somewhat faster Macintosh Quadra 700 (25 MHz),
but it still took several minutes to calculate a few hundred grains of sound.
Working at Les Ateliers UPIC in 1995, John Alexander and I developed the
Cloud Generator program (Roads and Alexander 1995). Cloud Generator is
a stand-alone synthesis and granulation program for MacOS computers. The
Appendix documents this program. Our implementation of Cloud Generator
merged the C code from several of my previous programs (Synthulate, Granulate, etc.) into a single interactive application. Since then, Cloud Generator has
served as a teaching aid in the basics of granular synthesis. It has also been used
in compositions by many musicians around the world. It provides a variety of
options for synthesis and sound processing. I have used it extensively for research purposes, and in composition.
Although Synthulate and its cousins have no graphical interface, they are
extensible. For this reason, I have continued to use them when I needed to try
an experiment that could not be realized in Cloud Generator. In early 1999, I
revised and recompiled Synthulate and its cousins for the Metrowerks C compiler on the Apple Power Macintosh computer.
Between 1996 and 2000, my CREATE colleagues and I also implemented a
variety of particle synthesis and sound processing programs using versions 1
and 2 of the SuperCollider language (McCartney 1996, 1998). SuperCollider
provides an integrated environment for synthesis and audio signal processing,
with gestural, graphical envelope, or algorithmic control. SuperCollider is my
synthesis environment of choice at the present time.

112

Chapter 3

Other Implementations of Granular Synthesis

The number of implementations of granular synthesis has increased greatly
in recent years. They are happening all over the world, running on dierent
platforms. Although I try here to be comprehensive, this survey is inevitably
incomplete.
Working at the Oberlin Conservatory, Gary Nelson carried out an experiment with something similar to granular synthesis in 1974. Beyond a brief
mention in (Nelson 1997), there seems to be no further documentation.
According to Clarke (1996), Michael Hinton implemented a type of granular
synthesis for a hybrid computer music system called IMPAC, at EMS Stockholm as early as 1984. When launched, the program generated a sequence of
short notes with pseudorandom variations on a number of parameters. These
included the upper and lower boundaries of a frequency region within which
the program scattered the notes. The synthesis was carried out by a combination of custom-made digital frequency modulation oscillators and analog
oscillators. A user could control various parameters in real time with either a
joystick or a digital pen. By 1988, Clarke had implemented FOF synthesis (a
granular technique; see chapter 4), on Atari computers and within the Csound
language (Clarke 1996). (See also Clarke 1998.)
The Canadian composer Barry Truax developed a series of important
implementations of granular synthesis. In 1986, he wrote a real-time application on the Digital Music Systems DMX-1000 signal processor, controlled by a
DEC LSI-11 microcomputer (Truax 1986). By 1987, he had modied his software to granulate a brief sampled sound. He achieved a technical breathrough
in 1990, making it possible to perform real-time granulation on an incoming
sound source, such as the live sound of an instrumentalist. This technique enabled him to realize a number of pioneering compositions (see chapter 7).
Truax later worked with engineers to develop the Quintessence Box for realtime granular synthesis, using a Motorola DSP 56001 chip for signal processing. A prototype of this box was demonstrated at the 1991 International Computer Music Conference. An operational unit was installed in 1993 at the
studios of Simon Fraser University, where the composer teaches.
Working at the University of Naples Federico II, Cavaliere, Evangelista,
and Piccialli (1988) constructed a circuit called the PSO Troll that could realize
up to sixteen voices of granular synthesis in real time at sampling rates up to
62.5 kHz.

113

Granular Synthesis

In the early 1990s, the Marseilles team of Daniel Arb and Nathalie Delprat
created the program Sound Mutations for time-frequency analysis of sound.
After analyzing a sound, the program modied and resynthesized it using
granular techniques. It could also perform transformations including timestretching, transposition, and ltering (Arb and Delprat 1992, 1993).
James McCartney included a granular instrument in his Synth-O-Matic program for MacOS (McCartney 1990, 1994). Users could draw envelopes on the
screen of the computer to control synthesis parameters.
Mara Helmuth realized two dierent implementations of granular synthesis techniques. StochGran was a graphical interface to a Cmix instrument
(Helmuth 1991). StochGran was originally developed for NeXT computers,
and later ported to the Silicon Graphics Incorporated IRIX operating system.
Helmuth also developed Max patches for granular sampling in real time on
the IRCAM Signal Processing Workstation (Helmuth 1993).
A group at the University of York implemented granular synthesis with
graphical control (Orton, Hunt, and Kirk 1991). A novel feature was the use of
cellular automata to modify the output by mapping the automata to the
tendency masks produced by the drawing program. Csound carried out the
synthesis.
In 1992 and 1993, I presented several lectures at IRCAM on granular synthesis and convolution techniques, After I left the institute, a number of people
who had attended these lectures launched granular synthesis and convolution
research of their own as extensions of other long-standing projects, namely
Chant synthesis and Max on the IRCAM Musical Workstation. The Granular
Synthesis Toolkit (GIST) consisted of a set of external objects for the Max
programming language, including a sinusoidal FOF grain generator, and a
FOG object for granulation (Eckel, Rocha-Iturbide, and Becker 1995; Rocha
1999). (See the description of FOF synthesis in chapter 4, and the description
of granulation in chapter 5.) Also at IRCAM, Cort Lippe (1993), developed
another Max application for granulation of sound les and live sound.
Recent versions of the Csound synthesis language (Boulanger 2000) provide four unit generators for granular synthesis: fof, fof2, grain, and granule.
Another unit generator, fog, was implemented in versions of Csound from
the universities of Bath and Montreal. The fof generator reads a synthetic
waveform function table and is oriented toward generating formant tones. The
fof2 generator adds control over the initial phase increment in the waveform
function table. This means that one can use a recorded sound and perform

114

Chapter 3

time-stretching, or extract segments. The grain unit generator begins reading a

waveform function table from a random point. The granule unit generator
handles up to four dierent grain streams with individual pitches. However,
most parameters (including a time-stretch factor) must be set at initialization
time (the beginning of each note in Csound), so the only parameter that can
be controlled during performance is grain density. The fog generator extracts
grains from a sound le. Lee (1995) also implemented a granular unit generator
for the Csound language.
The CDP GrainMill granular synthesis program runs on Windows. It derives
from Trevor Wishart's command line program Granula, a part of the Composer's Desktop Project System since 1996. The parameters aect each grain
individually as it is created. The parameters include size of the grain, density
control, time expansion and compression, pitch placement, amplitude, the portion of soundle from which the grain is extracted, spatial placement, and time
placement. The envelope of the grains is variable.
Tom Erbe's Cornbucket (1995) generated a granular synthesis score for
Csound. It oered envelope control for all synthesis parameters and was distributed in the form of source code in the C language.
Ross Bencina's Audiomulch is an audio signal processing application also for
Windows (Bencina 2000). It includes two modules for granulation of sampled
sounds. The Barcelona group of Lopez, Mart, and Resina (1998) developed
real-time granular synthesis (again, for Windows), featuring envelope, fader,
and MIDI control.
In 1995, R. De Tintis presented a paper to the Italian Computer Music
Association (AIMI) on his implementation of granular synthesis and sampling
on the IRIS-MARS workstation. The same year, Schnell (1995) and Todoro
(1995) implemented variants of granular synthesis on the IRCAM Musical
Workstation.
Kyma (Windows or MacOS) is a commercial package that oers real-time
granular synthesis. It is a graphical sound design environment in which one
interconnects graphical modules to construct synthesis patches. A synthesizer
called the Capybara renders the sound. The 1997 version of Kyma included
modules for granular synthesis, granular time expansion, and granular frequency scaling. The parameters of the grains (frequency, pitch deviation, rate of
emission, deviation in emission rate, waveform, grain envelope) are controllable
in real time through MIDI continuous controllers or faders displayed on the
screen, which also allow a source signal to be time-stretched and frequencyscaled.

115

Granular Synthesis

SuperCollider 2 (McCartney 1996, 1998; De Campo 1999) is a powerful

software environment for real-time audio synthesis that runs on MacOS computers. The SuperCollider 2 programming language oers an object-oriented
class system, a graphical interface builder for creating a patch control panel, a
graphical interface for creating wavetables and breakpoint envelopes, MIDI
control, a library of signal processing and synthesis functions, and a library of
functions for list processing of musical data. Users can write both the synthesis
and compositional algorithms for their pieces in the same high level language.
This allows the creation of synthesis instruments with considerably more exibility than is possible in other synthesis languages. SuperCollider can read and
write audio in real time or stream audio to or from a le. The new version,
SuperCollider 3, optimizes and extends these capabilities.
Gerhard Behles's real-time Granular program (Technical University Berlin)
runs on Silicon Graphics computers. The program reads a sound le and
manipulates it in real time. The user moves onscreen faders to change the eects
settings. The same author's Stampede allows composers to explore a continuum
of sound transformations under MIDI control. It performs granulation operations similar to those in Cloud Generator, but operates in real time. Andre
Bartetzki, at the electronic music studio of the Hochschule fur Musik Berlin,
has written a granular event generator called CMask that generates grain
specications for Csound (Bartetzki 1997a, 1997b). CMask provides numerous
options for scattering grains according to probabilistic functions, sieves, and
analogies to simple physical processes.
Damian Keller and Barry Truax (1998) developed Cmask models for
bouncing, breaking, scraping, and lling. The Cmask control functions determine where to scatter the grains in time and frequency. For example, a recursive equation approximated a bouncing pattern. By changing a damping
parameter, one could obtain a family of exponential curves with dierent rates
of damping or grain rate acceleration. Starting from samples of water drops,
Keller and Truax developed models of droplet patterns and streams, allowing
for a smooth transition between discrete droplets and denser aqueous sounds.
Chris Rolfe and Damian Keller (1999) developed a standalone MacOS program for soundle granulation called MacPOD.
William Mauchly of the company Waveboy created a granular synthesis
plugin for the Ensoniq ASR-10 and EPS-16 Plus samplers. Working as a signal
processor, users can granulate any sampled sound or live audio input. This
software oers time-scrambling, pitch-shifting, and adjustment of grain duration. Any MIDI controller can modulate the granulation parameters.

116

Chapter 3

Michael Norris (1997) provided four granulation processes in his SoundMagicFX package, which works with the SoundMaker program for MacOS.
Entitled Brassage Time Stretch, Chunk Munger, Granular Synthesis, and
Sample Hose, these exible procedures allow multiple-le input, time-varying
parameters, and additional signal processing to be applied to soundles, resulting in a wide range of granular textures.
Eduardo Miranda developed a Windows application called ChaosSynth for
granular synthesis using cellular automata (CA) control functions (Miranda
1998). Depending on how the CA are congured, they calculate the details of
the grains. A diculty posed by this approach is the conceptual rift between the
CA controls (number of cell values, resistances of the potential divider, capacitance of the electrical capacitor, dimension of the grid, etc.) and the acoustical
results (Correa, Miranda, and Wright 2000).
In 1999, Arboretum Systems oered a scattering granulator eect in its
popular Hyperprism eects processing software. The user controls grain size,
randomization, speed, as well as density and spread.
Can a standard MIDI synthesizer realize granular synthesis? Yes, in a limited
form. The New Yorkbased composer Earl Howard has done so on a Kurzweil
K2500 sampling synthesizer. The K2500 lets one create short samples, which
can repeat by internal signals as fast as 999 bpm, or about every 10 ms. Howard
created granular textures by layering several streams operating at dierent
rates, with each stream having a random delay. Another MIDI-based approach
to granular synthesis is found in Clarence Barlow's spectastics (spectral stochastics) technique. This generates up to two hundred notes per second to approximate the spectrum of a vocal utterance (Barlow 1997).
Even with all these implementations, there is still a need for an instrument
optimized with controllers for the virtuoso performance of granular textures.
Apropos of this, see the description of the Creatovox project in chapter 5.

Summary
As regards electric instruments for producing sound, the enmity with which the few musicians who know them is manifest. They judge them supercially, consider them ugly, of
small practical value, unnecessary. . . . [Meanwhile, the inventors] undiscerningly want the
new electric instruments to imitate the instruments now in use as faithfully as possible and
to serve the music that we already have. What is needed is an understanding of the . . .
possibilities of the new instruments. We must clearly evaluate the increase they bring to our

117

Granular Synthesis
own capacity for expression . . . The new instruments will produce an unforeseen music, as
unlooked for as the instruments themselves. (Chavez 1936)

Granular synthesis is a proven method of musical sound synthesis, and is featured in important compositions (see chapter 7). Implementations of granular
techniques are widespread. Most focus on the granulation of sampled sound
les. Pure granular synthesis using synthetic waveforms is available only in a
few packages.
At low densities, synchronous GS serves as a generator of metrical rhythms
and precise accelerandi/rallentandi. A high-density cloud set to a single frequency produces a stream of overlapping grains. This forms sweet pitched tones
with strong formants, whose position and strength depend greatly on the grain
envelope and duration.
Asynchronous GS sprays thousands of sonic grains into cloudlike formations
across the audio spectrum. At high densities the result is a scintillating sound
complex that varies over time. In musical contexts, these types of sounds can
act as a foil to the smoother, more sterile sounds emitted by digital oscillators.
Granulation of sampled sounda popular techniqueproduces a wide range
of extraordinary variations, explored in chapter 5. The destiny of granular
synthesis is linked both to graphics and to real-time performance.
A paint program oers a uid interface for granular synthesis. The MetaSynth program (Wenger and Spiegel 1999), for example, provides a spray brush
with a variable grain size. A further extension would be a multicolored spray jet
for sonic particles, where the color palette corresponds to a collection of waveform samples. (In MetaSynth, the color of the grains indicates their spatial
location.)
Analysis/resynthesis systems, such as the phase vocoder, have an internal
granular representation that is usually hidden from the user. As predicted (in
Roads 1996), the interfaces of analysis/resynthesis systemswhich resemble
sonogramshave merged with interactive graphics techniques. This merger
sonographic synthesisis a direct and intuitive approach to sound sculpture.
(See chapters 4 and 6 for more on sonographic synthesis and transformation.)
One can scan a sound image (sonogram), touch it up, paint a new image, or
erase it, with the algorithmic brushes of computer graphics.
My colleagues and I continue to rene our instrument for real-time virtuoso
performance of granular synthesis (Roads 19921997). The Creatovox research
project at the University of California, Santa Barbara has resulted in a prototype of a granular synthesis instrument, playable on a standard musical keyboard and other controllers. (See the description in chapter 5.)

118

Chapter 3

Granular synthesis oers unique opportunities to the composer and suggests new ways of organizing musical structureas clouds of evolving sound
spectra. Indeed, granular representation seems ideal for representing statistical
processes of timbral evolution. Time-varying combinations of clouds lead to
such dramatic eects as evaporation, coalescence, and mutations created by crossfading overlapping clouds. A striking similarity exists between these processes
and those created in computer graphics by particle synthesis (Reeves 1983),
often used to create images of re, water, clouds, fog, and grasslike textures,
analogous to some of the audio eects possible with asynchronous granular
synthesis.

Varieties of Particle Synthesis

Glisson Synthesis
Magnetization Patterns of Glisson Clouds
Implementation of Glisson Synthesis
Experiments with Glisson Synthesis
Assessment of Glisson Synthesis
Grainlet Synthesis
Parameter Linkage in Grainlet Synthesis
Frequency-Duration Experiments
Amplitude-Duration Experiments
Space-Duration Eperiments
Frequency-Space Experiments
Amplitude-Space Experiments
Assessment of Grainlet Synthesis
Trainlet Synthesis
Impulse Generation
Theory and Practice of Trainlets
Assessment of Trainlet Cloud Synthesis
Pulsar Synthesis
Basic Pulsar Synthesis
Pulsaret-Width Modulation
Synthesis across Time Scales
Spectra of Basic Pulsar Synthesis

120

Chapter 4

Advanced Pulsar Synthesis

Multiple Pulsar Generators
Pulse Masking, Subharmonics, and Long Tonepulses
Convolution of Pulsars with Samples
Implementations of Pulsar Synthesis
Composing with Pulsars
Musical Applications of Pulsar Synthesis
Assessment of Pulsar Synthesis
Graphic and Sonographic Synthesis of Microsound
Graphic Synthesis in the Time-Domain and Frequency-Domain
Micro-Arc Synthesis with the UPIC
Synthesis of Microsound in Phonogramme
Synthesis of Microsound in MetaSynth
Assessment of Graphic and Sonographic Synthesis of Microsound
Particle-Based Formant Synthesis
FOF Synthesis
Vosim
Window Function Synthesis
Assessment of Particle-Based Formant Synthesis Techniques
Synthesis by Transient Drawing
Assessment of Transient Drawing
Particle Cloning Synthesis
Assessment of Particle Cloning Synthesis
Physical Models of Particles
Assessment of Physical Models of Particles
Abstract Models of Particles
Abstract Particle Synthesis
Assessment of Abstract Models of Particles
Summary

121

Varieties of Particle Synthesis

There are two kinds of experimentalists. One kind talks to theorists. The theorist makes
predictions and the experimentalist then does the experiments. To do this is important, but
then all you do is follow the theory. Another type designs his own experiments, and in this
way is ahead of the theorists.
Samuel C. C. Ting (Nobel Prize in Physics 1988)

Unlike particles probed by physicists, synthetic particles inhabit a virtual world.

This world is invented, and the waveforms produced by the particles are
algorithmically derived. They may be simple and regular in structure, forming smooth pitched tones, or complex and irregular, forming crackling noisy
masses.
The engines of particle synthesis are not especially complicated. It is the
combination of many simple elements that forms a complex time-varying
sound. We shape the sound's evolution by controlling this combination from a
high musical level. High-level controls imply the existence of algorithms that
can interpret a composer's directives, translating them into low-level particle
specications. (See chapter 8's discussion of simplicity versus complexity in
microsound synthesis.)
This chapter presents a catalog of particle synthesis techniques. These include
glissons, grainlets, trainlets, pulsars, graphic and sonographic particles, formant
particles, transient drawing, particle cloning, and physical and abstract models
of particles.

Glisson Synthesis
Glisson synthesis is an experimental technique of particle synthesis. It derives
from the technique of granular synthesis, presented in the previous chapter. I
implemented glisson synthesis after revisiting Iannis Xenakis's original paper
on the theory of granular synthesis (Xenakis 1960). In this article, Xenakis
described each grain as a vector within a three-dimensional space bounded by
time, frequency, and amplitude. Since the grain is a vector, not a point, it can
vary in frequency, creating a short glissando. Such a signal is called a chirp or
chirplet in digital signal processing (Mann and Haykin 1991). Jones and Parks
implemented frequency-modulated grains with a variable chirp rate in 1988.
My implementation of glisson synthesis dates to 1998.
In glisson synthesis, each particle or glisson has an independent frequency
trajectoryan ascending or descending glissando. As in classic granular synthesis, glisson synthesis scatters particles within cloud regions inscribed on the

122

Chapter 4

time-frequency plane. These clouds may be synchronous (metric) or asynchronous (ametric). Certain parameters of glisson synthesis are the same as for
granular synthesis: start time and duration of the cloud, particle duration,
density of particles per second, frequency band of the cloud, amplitude envelope of the cloud, waveform(s) within the particles, and spatial dispersion of the
cloud. (See the description in the previous chapter.)
Magnetization Patterns of Glisson Clouds
The magnetization patterna combination of several parametersdetermines
the frequency direction of the glissons within a cloud. First, the glissandi may
be deep (wide frequency range) or shallow (small frequency range) (gure 4.1a,
b). Second, they may be unidirectional (uniformly up or down) or bidirectional
(randomly up or down) (gure 4.1c, d, e). Third, they may be diverging (starting from a common center frequency and diverging to other frequencies), or
converging (starting from divergent frequencies that converge to a common
center frequency). The center frequency can be changing over time.
Implementations of Glisson Synthesis
Stephen Pope and I developed the rst implementation of glisson synthesis in
February 1998. The software was coded in the SuperCollider 1 synthesis language (McCartney 1996, Pope 1997). Later, I modied the glisson program and
carried out systematic tests. In the summer of 1999, Alberto de Campo and I
reimplemented glisson synthesis in the SuperCollider 2 language (McCartney
1998).
Experiments with Glisson Synthesis
Short glissons (< 10 ms) with a large frequency variation (> 100 Hz) resemble
the classic chirp signals of digital signal processing; sweeping over a wide frequency range in a short period of time. An individual glisson of this type in the
starting frequency range of 400 Hz sounds like a tap on a wood block. When
the starting frequency range is around 1500 Hz, the glissons sound more like
the tapping of claves. As the density of glissons increases and the deviation
randomizes in direction, the texture tends quickly toward colored noise.
Medium-length (25100 ms) glissons ``tweet'' (gure 4.2a), so that a series of
them sounds like birdsong.

123

Varieties of Particle Synthesis

Figure 4.1 Magnetization patterns in glisson synthesis. The vertical axis is frequency
and the horizontal axis is time. (a) Shallow (small frequency deviation) bidirectional.
(b) Deep (large frequency deviation) bidirectional. (c) Upwards unidirectional. (d)
Downwards unidirectional. (e) Diverging from center frequency. (f ) Converging to
center frequency.

Long glissons (> 200 ms) result in dramatic cascades of sound (gure 4.2b).
At certain densities, they are reminiscent of the massed glissandi textures heard
in such orchestral compositions as Xenakis's Metastasis (1954). A striking eect
occurs when the glissandi diverge from or converge upon a common central
frequency. By constraining the glissandi to octaves, for example, it is possible to
generate sounds similar to the Shepard tones (Risset 1989a, 1997), which seem
to spiral endlessly upward or downward.
Assessment of Glisson Synthesis
Glisson synthesis is a variant of granular synthesis. Its eects segregate into two
categories. At low particle densities, we can perceive each glissando as a sepa-

124

Chapter 4

Figure 4.2 Glissons. (a) Sonogram of a single 25-ms glisson. Notice the gray artefacts
of the analysis, reecting the time-frequency uncertainty at the beginning and end of the
particle. (b) Glisson cloud generated by a real-time performance. The glisson durations
increase over the 16-second duration of the cloud.

125

Varieties of Particle Synthesis

rate event in a micro-melismatic chain. When the glissons are short in duration
(< 50 ms), their internal frequency variation makes it dicult to determine
their pitch. Under certain conditionssuch as higher particle densities with
greater particle overlapglisson synthesis produces second-order eects that
we perceive on the time scale of sound objects. In this case, the results tend toward a mass of colored noise, where the bandwidth of the noise is proportional
to the frequency variation of the glissandi. Several factors can contribute to the
sensation of a noise mass, the most important being density, wide frequency
variations, and short glisson durations.

Grainlet Synthesis
Grainlet synthesis combines the idea of granular synthesis with that of wavelet
synthesis. (See chapter 6.) In granular synthesis, the duration of a grain is
unrelated to the frequency of its component waveform. In contrast, the wavelet
representation scales the duration of each particle according to its frequency.
Short wavelets represent high frequencies, and long wavelets represent low
frequencies. Grainlet synthesis generalizes this linkage between synthesis
parameters. The fundamental notion of grainlet synthesis is that any parameter of synthesis can be made dependent on (or linked to) any other parameter.
One is not, for example, limited to an interdependence between frequency and
duration.
I implemented grainlet synthesis in 1996 as an experiment in parameter
linkage within the context of granular cloud synthesis (described in the previous chapter). Grainlet synthesis imposes no constraints on the choice of waveform, particle envelope, or any other parameter, except those that we introduce
through parameter linkage.
Parameter Linkage in Grainlet Synthesis
Parameter linkage is the connecting of one parameter with a dependent parameter. As parameter A increases, for example, so does its dependent parameter B. One can also stipulate inverse linkages, so that an increase in A results in
a decrease in B.
Parameter linkages can be drawn as patch diagrams connecting one parameter to another (gure 4.3). An arrow indicates a direct inuence, and a gray

126

Chapter 4

Figure 4.3 Parameter linkage in grainlet synthesis. Each circle represents a parameter
of grainlet synthesis. An arrow from one parameter to another indicates a dependency.
Here parameter 7 is dependent on parameter 2. If parameter 7 is spatial depth and parameter 2 is grainlet start time, then later grainlets have more reverberation. Notice that
parameter 4 is inversely dependent on parameter 8, as indicated by the gray dot. If parameter 4 was grainlet duration and parameter 8 was grainlet frequency, then higher
frequency grainlets are shorter in duration (as in wavelet resynthesis).

circle an inverse linkage. I wrote a program in the C language to realize these

parameter linkages. In the rst version, grainlet duration could be specied
in terms of the number of cycles of the fundamental period. If the number of
cycles is ten, for example, a grainlet at 100 Hz lasts 10  0:01 sec 0:1 sec,
while a grainlet at 1000 Hz lasts 10  0:001 Hz 0:01 sec.
After initial tests, I generalized the parameter linkage from frequency and
duration to dependencies between any two synthesis variables. The grainlet
synthesis program generates a data le. A Csound program for granular synthesis interprets this le and synthesizes the sound. The synthesis parameters
include the following:
1 Cloud density (number of grainlets per second)
1 Grainlet amplitude
1 Grainlet start-time
1 Grainlet frequency
1 Grainlet duration

127

Varieties of Particle Synthesis

Figure 4.4 Collections of grainlets. (a) These grainlets are scaled in duration according
to their frequency. (b) Superposition of short high-frequency grainlets over a long lowfrequency grainlet.

1 Grainlet waveform
1 Grainlet position in the stereo eld
1 Grainlet spatial depth (amount of reverberation)
Frequency-Duration Experiments
The rst experiments with grainlet synthesis simulated the relationship between
grain duration and grain frequency found in wavelet representation (gure 4.4a
and b). I later generalized this to allow any frequency to serve as a point of attraction around which certain durations (either very long or very short) could
gravitate (gure 4.5).

128

Chapter 4

Figure 4.5 Inverse sonogram plotted on a logarithmic frequency scale, showing a frequency point of attraction around the grainlet spectrum. The grainlets whose frequencies
are close to the point of attraction (700 Hz) are long in duration, creating a continuous
band centered at this point.

Amplitude-Duration Experiments
These experiments linked grain duration with the amplitude of the grains. In
the case of a direct link, long grains resulted in louder grains. In an inverse relationship, shorter grains had higher amplitudes.
Space-Duration Experiments
These experiments positioned grains in space according to their duration.
Grains of a stipulated duration always appeared to emanate from a specic
location, which might be any point in the stereo eld. Grains whose duration
was not stipulated scattered randomly in space.
Frequency-Space Experiments
These experiments positioned grains in space according to their frequency.
Grains of a stipulated frequency appeared to always emanate from a specic
location, which might be any point in the stereo eld. Other grains whose frequency was not stipulated scattered randomly in space.

129

Varieties of Particle Synthesis

Amplitude-Space Experiments
These experiments assigned grains a spatial location according to their amplitude. Grains of a stipulated amplitude appeared to always emanate from a
specic location, which might be any point in the stereo eld. Other grains
whose amplitude was not stipulated scattered randomly in space.
Assessment of Grainlet Synthesis
Grainlet synthesis is an experimental technique for realizing linkages among the
parameters of microsonic synthesis. It appears to be a good technique for forcing high-level organizations to emerge from microstructure. Specically, the
clouds generated by grainlet synthesis stratify, due to the internal constraints
imposed by the parameter linkages. This stratication is seen in textures such
as a dense cloud of brief, high-frequency grains punctuated by low and long
grains. Other clouds stratify by spatial divisions. Many parameter linkages are
easy to discern, conveniently serving as articulators in music composition.

Trainlet Synthesis
A trainlet is an acoustic particle consisting of a brief series or train of impulses.
Like other particles, trainlets usually last between 1 to 100 ms. To create timevarying tones and textures, an algorithm is needed that can spawn thousands of
trainlets from a few high-level specications. The main parameters of trainlet
synthesis are the density of the trainlets, their attack time, pulse period, harmonic structure, and spectral energy prole. Before explaining the theory of
trainlets, let us summarize the basics of impulse generation.
Impulse Generation
An impulse is an almost instantaneous burst of energy followed by an immediate decline in energy. In its ideal form, an impulse is innitely narrow in the
time dimension, creating a single vertical line in its time-domain prole. In
practice, however, impulses always last a nite time; this is their pulse width.
Electronic impulses in the real world vary greatly, exhibiting all manner of attack shapes, decay shapes, and transition times. These variations only make
them more interesting from a musical point of view.

130

Chapter 4
Table 4.1 Technical specications of the Hewlett-Packard HP8005B pulse generator
Repetition rate
Attack and decay transition times
Overshoot, preshoot, and ringing
Pulse width
Width jitter
Pulse delay
Delay jitter
Period jitter

0.3 Hz to 20 mHz in ve ranges; Vernier control

within each range
<10 ns to 2 s in ve ranges; Verniers for leading
and trailing edges
<5% of pulse amplitude each
<25 ns to 3 s in ve ranges; Vernier controls
within each range
<0.1% 50 ps of any width setting
<100 ns to 3 s, in ve ranges
<0.1% 50 ps of any delay setting
<0.1% of any period setting

Impulse generation is one of the most venerable methods of electronic sound

synthesis. At infrasonic frequencies, we perceive impulses separately and notice
their rhythmic pattern. At audio frequencies the impulses fuse into continuous
tones. Electronic music studios of the 1950s were equipped with laboratory
impulse generators (IGs). The IGs installed in the WDR Cologne studio had a
pulse frequency range that extended from the infrasonic frequencies (1.1 Hz,
corresponding to a pulse duration of 900 ms) to the high audio frequencies
(10 kHz, or pulse duration of 0.1 ms).
Laboratory impulse generators have evolved considerably since the 1950s.
Modern IGs are solid-state analog devices. Table 4.1 lists the specications of a
low-voltage IG in the price range of $600. Notice the wide frequency range,
from 0.15 Hz (one impulse every 6.6 seconds) to 20 MHz. The spectrum of an
impulse with a period of 50 nanoseconds (corresponding to a frequency of
20 MHz) is very broad. More expensive IGs, costing up to $25,000, extend the
fundamental frequency range beyond 3 gHz. The main parameters of an IG are
the shape of the impulse, the fundamental period of repetition, and the duty
cycle of the pulse wave relative to the period. These parameters aect the
rhythm or pitch of the output signal, and its overall spectrum.
Industrial impulse generators are not designed for exible performance in
real time. Notice the specications in table 4.1, where repetition rate, transition
time, pulse width, and pulse delay are broken into ve or six ranges. This means
that they cannot sweep continuously from one setting to another. They are also
not designed for automated operation.
Moreover, the raw output of an IG is a series of clicks, not sucient in
themselves for music synthesis. Since a click has a broad bandwidth, a common

131

Varieties of Particle Synthesis

synthesis technique is to lter the impulses with a bandpass characteristic to

carve out a specic peak in the spectrum. When the lter is very narrow, a single impulse causes the lter to resonate at a specic frequency, yielding the
sensation of pitch. Another processing trick is to pass the impulse train through
a variable delay and feedback loop, so that the density of impulses accumulates
under the control of the composer. Karlheinz Stockhausen and Gottfried
Michael Koenig used these kinds of techniques in the synthesis of Kontakte in
1960. (See also Kaegi 1967.)
Ultimately, all electronic signals, whether analog or digital, are bandlimited.
An analog IG has an obvious advantage over a digital IG, which is strictly
bandlimited by the Nyquist theorem. As is clear from the specications in table
4.1, the bandwidth of analog IGs is much broader.
Theory and Practice of Trainlets
A trainlet cloud contains one or more trainlets. The main variables in the synthesis of a trainlet cloud are:
1 Density per second of trainlets
1 Trainlet pulse period
1 Trainlet harmonic structure
1 Trainlet spectral energy prole
An amplitude envelope shapes each trainlet, and the generator projects each
trainlet to a specic location in space. The parameters vary for each trainlet, so
each trainlet may be unique.
Since trainlet synthesis operates in the digital domain, it is important that the
impulses be bandlimited to avoid aliasing. A standard bandlimited impulse
generator for modern synthesis languages is Godfrey Winham's buzz unit generator and its variant, gbuzz (Winham 1966; Winham and Steiglitz 1970; Howe
1975; Steiglitz 1996). Buzz emits a set of N harmonic partials of equal amplitude. As the number of harmonics N increases, the pulse width becomes narrower (gure 4.6). Gbuzz emits a set of N harmonic partials with more exible
control of the spectrum.
In the 1998 prototype implementation of trainlet synthesis, I wrote the synthesis engine in the Csound language (Boulanger 2000). Csound's gbuzz unit
generator emits harmonic cosines aligned to a fundamental frequency. It scales
the amplitudes of the cosines so their sum equals a stipulated peak value. One

132

Chapter 4

Figure 4.6 Bandlimited pulses. (a) Sum of eight harmonics. (b) Sum of thirty-two harmonics. Notice the narrowing of the pulse.

can stipulate the lowest harmonic, the total number of harmonics starting from
the lowest, and the chroma of the harmonic series (see below).
Chroma is a spectral brightness factor. Figure 4.7 shows the relationship
between chroma and spectra. Chroma determines the relative strength of the
harmonic series. If the lowest harmonic partial has a strength coecient of A,
the lowest harmonic nth partial will have a coecient of A  chroma n , an
exponential curve. The chroma may be positive, zero, or negative, and is not
restricted to integers. If chroma 1, the harmonics are of equal strength. If
chroma < 1, the higher harmonics are attenuated, as though the signal had
been sent through a lowpass lter. As the value of chroma tends toward 0, they
attenuate more rapidly. If chroma > 1, the highest harmonic has the greatest
amplitude (as though a highpass lter had processed it), while each lower harmonic stepping down from the highest has a progressively lower amplitude. As
the chroma value increases, the signal is brighter in timbre.

133

Varieties of Particle Synthesis

Figure 4.7 Relationship between chroma and spectra. We see time-domain waveforms
and sonogram spectra (on a logarithmic frequency scale) of trainlets with increasing
chroma. All trainlets have a pulse frequency of 100 Hz, with thirty-two harmonics, and
the lowest harmonic is always 1. The chroma values are indicated at the bottom of the
gure. In the last example, chroma 20, the algorithm explodes.

134

Chapter 4

Figure 4.8 Time-domain view of scattering of trainlets in two channels. The trainlets
are at various frequencies and have various durations. The amplitude of each trainlet
(the sum of both channels) is constant.

Trainlet clouds are musical units on the sound object time scale. As in granular synthesis, synchronous clouds spawn a regular series of trainlets. A linear
accelerando or rallentando can also be realized in this mode. Asynchronous
clouds spawn trainlets at random intervals according to the stipulated density.
Figure 4.8 provides a time-domain view of a trainlet cloud, while gure 4.9
shows the sonogram of two trainlet clouds, one with long trainlets, the other
with short trainlets. Notice their characteric spectral pattern.
Table 4.2 enumerates the rest of the cloud parameters. The values in brackets
are typical ranges.

135

Varieties of Particle Synthesis

Figure 4.9

Sonogram of trainlet clouds. (a) Long trainlets. (b) Short trainlets.

136

Chapter 4
Table 4.2 Trainlet cloud parameters
1. Cloud start time, in seconds
2. Cloud duration, in seconds
3. Random duration ag. If set, interpret the initial and nal trainlet durations as
maximum and minimum trainlet durations, respectively. The actual durations are
generated randomly between these limits.
4. Trainlet durations at start of cloud (0.0010.8 sec)
5. Trainlet durations at end of cloud
6. Density of trainlets per second at start of cloud (1300)
7. Density of trainlets per second at end of cloud
8. Upper frequency bandlimit of the cloud at start (20 Hz20 KHz)
9. Lower frequency bandlimit of the cloud at start
10. Upper frequency bandlimit of the cloud at end
11. Lower frequency bandlimit of the cloud at end
12. Amplitude at start of cloud (196 dB)
13. Amplitude at end of cloud
14. Number of harmonics per trainlet at start of cloud (164)
15. Number of harmonics per trainlet at end of cloud
16. Lowest sounding harmonic at start of cloud
17. Lowest sounding harmonic at end of cloud
18. Chroma at start of cloud. If chroma < 0 then the eect is lowpass. If chroma 1
all harmonics are equal in strength. If chroma > 1 then the eect is highpass.
19. Chroma at end of cloud,
20. Initial waveform, usually sine
21. Final waveform, usually sine
22. Spatial position of trainlets in the stereo eld at the start of the cloud, either xed
(0 L, 0:5 middle, 1 right) or random
23. Spatial position of trainlets in the stereo eld at the end of the cloud
24. Initial attack time (550 ms)
25. Final attack time
Note: Initial and nal values refer to the beginning of the cloud and the end of the cloud,
respectively.

137

Varieties of Particle Synthesis

We have also implemented a version of trainlet synthesis in the real-time

Creatovox synthesizer, described in chapter 5. One hand can play the trainlet
clouds on the keyboard of the instrument, while the other manipulates the
cloud parameters with joysticks and other MIDI controllers.
Assessment of Trainlet Cloud Synthesis
I conducted numerous tests with the trainlet cloud technique focussing primarily on trainlets with exponential attacks, linear sustains, and linearly decaying
envelopes. The gbuzz unit generator, stamps the trainlet sonority with its
``brittle'' signature (Schindler 1998). This brittleness derives from gbuzz's xed
waveform and unnatural spectrum. Long trainlets in sparse clouds revisit a
sound quality that was characteristic of early computer music studies.
My best results involved trainlets with durations of 40 ms to 100 ms, where
the characteristic timbre of the trainlets had just enough time to manifest.
Fundamental frequencies greater than 500 Hz tend to sound metallic, fundamentals under 500 Hz are darker and more eective. In any case, with trainlets
having up to thirty-two harmonics, the fundamental must remain under 600 Hz
in order to avoid aliasing. A 32-harmonic trainlet at 600 Hz with a chroma of
1.0 is very bright.
With over twenty parameters to vary, a wide range of sound material can be
produced by trainlet cloud synthesis, which seems to have considerable musical
potential. More research is needed, however. As with any technique, the output
of trainlet synthesis can be taken as a starting point for further signal processing. I have, for example, tested congurations in which a trainlet generator
connects to a constant-Q bandpass lter. The lter lets one accentuate a particular formant in the partial spectrum.

Pulsar Synthesis
Pulsar synthesis (PS), named after the spinning neutron stars that emit periodic
signals in the range of 0.25 Hz to 642 Hz, is a powerful method of digital sound
synthesis with links to past analog techniques. Coincidentally, this range of
frequenciesbetween rhythm and toneis of central importance in pulsar
synthesis.
PS melds established principles within a new paradigm. In its basic form, it
generates electronic pulses and pitched tones similar to those produced by

138

Chapter 4

analog instruments such as the Ondioline (Jenny 1958; Fourier 1994) and the
Hohner Elektronium (1950), which were designed around the principle of
ltered pulse trains. Pioneering electronic music composers including Stockhausen (1955, 1957, 1961, 1963) and Koenig (1959, 1962) used ltered impulse
generation as a staple in their studio craft. Pulsar synthesis is a digital technique, however, and so it accrues the advantages of precise programmable
control, waveform exibility, graphical interface, and extensibility. In its advanced form, pulsar synthesis generates a world of rhythmically structured
crossbred sampled sounds.
This section rst presents the basic theory of pulsars and pulsar graphs. We
then move on to the more advanced technique of using pulsars to transform
sampled sounds through cross-synthesis, presenting musical applications of
pulsar synthesis in compositions by the author. At the end of this section, we
describe the features of a new interactive program called PulsarGenerator
(Roads 2001).
Basic Pulsar Synthesis
Basic pulsar synthesis generates a family of classic electronic music timbres akin
to those produced by an impulse generator connected to a bandpass lter.
Unlike this classic technique, however, there is no lter in the basic PS circuit.
A single pulsar is a particle of sound. It consists of an arbitrary pulsaret
waveform w with a period d followed by a silent time interval s (gure 4.10a).
The total duration of a pulsar is p d s, where p is the pulsar period, d is the
duty cycle, and s is silent. Repetitions of the pulsar signal form a pulsar train.
Let us dene the frequency corresponding to the repetition period as fp 1=p
and the frequency corresponding to the duty cycle as fd 1=d. Typical ranges
of fp are between 1 Hz and 5 kHz, the typical range of fd is from 80 Hz to
10 kHz.
In PS, both fp and fd are continuously variable quantities. They are controlled by separate envelope curves that span a train of pulsars. The train is the
unit of musical organization on the time scale of notes and phrases, and can last
anywhere from a few hundred milliseconds to a minute or more.
Notice in gure 4.10b that the duty ratio or d : s ratio varies while p remains
constant. In eect, one can simultaneously manipulate both fundamental frequency (the rate of pulsar emission) and what we could call a formant frequency
(corresponding to the duty cycle), each according to separate envelopes. A

139

Varieties of Particle Synthesis

Figure 4.10 Anatomy of a pulsar. (a) A pulsar consists of a brief burst of energy called
a pulsaret w of a duration d followed by a silent interval s. The waveform of the pulsaret,
here shown as a band-limited pulse, is arbitrary. It could also be a sine wave or a period
of a sampled sound. The total duration p d s, where p is the fundamental period
of the pulsar. (b) Evolution of a pulsar train, time-domain view. Over time, the pulsar
period p remains constant while the pulsaret period d shrinks. The ellipses indicate a
gradual transition period containing many pulsars between the three shown.

140

Chapter 4

Figure 4.11 Typical pulsaret waveforms. In practice, any waveform can be used.
(a) Sine. (b) Multicycle sine. (c) Band-limited pulse. (d) Decaying multicycle sinusoid.
(e) Cosmic pulsar waveform emitted by the neutron star Vela X-1.

formant is a peak region in a spectrum. Lowering the fundamental means

increasing s, and raising the fundamental means decreasing s.
So far, the structure we have described is similar to that of a standard impulse generator. Pulsar synthesis generalizes this conguration in several ways.
First, it allows the pulsaret w to be any waveform. Figure 4.11 shows some
typical pulsaret waveforms, including those with multiple subperiods within
their duty cycle (gure 4.11b, d, and e).
Let us assume that w is a single cycle of a sine wave. From a signal processing point of view, this can be seen as a sine wave that has been limited in time
by a rectangular function v, which we call the pulsaret envelope. An important

141

Varieties of Particle Synthesis

Figure 4.12 Typical pulsaret envelopes. (a) Rectangular. (b) Gaussian. (c) Linear
decay. (d) Exponential decay. The term b determines the steepness of the exponential
curve. (e) Linear attack, with duty cycle d. (f ) Exponential attack. The term x determines
the steepness of the exponential curve. (g) FOF envelope. (h) Bipolar modulator.

generalization is that v can also be any shape. As we show later, the envelope v
strongly aects the spectrum of the pulsar train.
Figure 4.12 shows some typical pulsaret envelopes. A rectangular envelope
(gure 4.12a) produces a broad spectrum with strong peaks and nulls for any
pulsaret. Figure 4.12g depicts a well-known conguration for formant synthesis, an envelope with a sharp attack followed by an exponential decay. This
corresponds to the FOF and Vosim techniques described later in this chapter.
Such a conguration can be seen as a special case of pulsar synthesis. As gure
4.12h shows, the envelope can also be a bipolar ring modulator.
Keeping p and w constant and varying d on a continuous basis creates the
eect of a resonant lter swept across a tone. There is, of course, no lter in this
circuit. Rather, the frequency corresponding to the duty cycle d appears in the
spectrum as a formant peak. By sweeping the frequency of this peak over time,
we obtain the sonic equivalent of a time-varying bandpass lter applied to a
basic impulse train.

142

Chapter 4

Pulsaret-Width Modulation
Pulse-width modulation (PWM) is a well-known analog synthesis eect occuring
when the duty cycle of a rectangular pulse varies while the fundamental frequency remains constant (gure 4.13a). This produces an edgy ``sawing'' quality as the upper odd harmonics increase and decrease over the course of the
modulation. At the extremes of PWM, the signal is silent. For example, when
d 0, PWM results in a signal of zero amplitude (gure 4.13b). When d p,
PWM produces a signal of a constant amplitude of 1 (gure 4.13c).
Pulsaret-width modulation (PulWM) extends and improves this model. First,
the pulsaret waveform can be any arbitrary waveform. Second, it allows the
duty cycle frequency to pass through and below the fundamental frequency.
Here fd a fp . Notice in gure 4.13 how the duty cycle of the sinusoid increases
from (d) to (e). In (f ), p d. Finally, in (g) p < d. That is, the duty cycle is
longer than the fundamental period. Only the rst quadrant of the sine wave
repeats. The fundamental period cuts o the duty cycle of the pulsaret in midwaveform. In our implementation, we apply a user-controlled crossfade time
around this cuto point, which we call the edge factor. When there is no
crossfade, the edge factor is high.
We have also tested an alternative approach to pulsar-width modulation,
designed by Alberto de Campo, which produces a dierent sound. In this
method, overlapped pulsaret-width modulation or OPulWM, the fundamental
frequency corresponds to the rate of pulsar emission, independent of the pulsaret duty cycle. That is, the duty cycle of an individual pulsar always completes, even when it crosses below the fundamental frequency. Whenever the
fundamental period expires, our algorithm spawns a new pulsar. Thus, when
d > p several pulsars overlap with others whose duty cycle has not yet completed. As d increases, the generator spawns more and more overlapping
pulsars. For practical reasons, then, we stipulate an arbitrary overlap limit.
OPulWM results in a great deal of phase cancellation and so tends to be a more
subtle eect than regular PulWM.
Synthesis across Time Scales
PS operates within and between musical time scales. It generates a stream of
microsonic particles at a variable rate across the continuum spanning infrasonic pulsations and audio frequencies. When the distance between successive

143

Varieties of Particle Synthesis

Figure 4.13 Pulsaret-width modulation. (a) Classical PWM with a rectangular pulse
shape. The ellipses indicate a gradual transition between the pulses. (b) PWM when the
duty cycle d 0 results in a signal of zero amplitude. (c) PWM when the duty cycle
d p (the fundamental period), the result is a signal with a constant amplitude of 1.
(d) Pulsar train with a sinusoidal pulsaret. (e) Same period as (d), but the duty cycle is
increasing. (f ) The duty cycle and the period are equal, resulting in a sinusoid. (g) The
duty cycle is greater than the fundamental period, which cuts o the nal part of the sine
waveform.

144

Chapter 4

impulses is less than about one twentieth of a second, the human hearing
mechanism causes them to fuse into a continuous tone. This is the forward
masking eect (Buser and Imbert 1992). As Helmholtz (1885) observed, in
the range between 20 and 35 Hz, it is dicult to distinguish the precise pitch
of a sustained tone; reliable pitch perception takes hold at about 40 Hz,
depending on the waveform. So listeners hear pitch in a periodic sustained tone
for p between approximately 25 ms (corresponding to fp 40 Hz) and 200 msec
(corresponding to fp 5 kHz).
As the rate of pulsar emission slows down and crosses through the threshold
of the infrasonic frequencies ( fp < 20 Hz), the sensation of continuous tone
evaporates, and we can perceive each pulsar separately. When the fundamental
fp falls between 62.5 ms (corresponding to the time span of a thirty-second note
at quarter note 60 MM) and 8 sec (corresponding to the time span of two
tied whole notes at quarter note 60 MM), we hear rhythm. The fundamental
frequency envelope becomes a graph of rhythm. This takes the form of a function of time that a user draws on the screen of a computer. Such a pulsar graph
can serve as an alternative form of notation for one dimension of rhythmic
structure, namely the onset time of events. The correspondence between the
musical units of rhythmic structure (note values, tuplets, rests, etc.) can be made
clear by plotting note values on the vertical or frequency scale. For example,
assuming a tempo of 60 MM, a frequency of 5 Hz corresponds to a quintuplet
gure. Note that the two-dimensional pulsar graph does not indicate the duration of the events. This could be represented by adding a third dimension to the
plot.
To interpret the rhythm generated by a function inscribed on a pulse graph,
one has to calculate the duration of the grain emission curve at a given xed
frequency rate. For example, a grain emission at 4 Hz lasting 0.75 seconds
emits three grains. When grain emission switches from one value to the next,
the pulsar corresponding to the new duration plays immediately followed by a
silence equal to the period of grain emission. Figure 4.14 plots a rhythm that
alternates between xed-rate pulses, accelerandi, and silence.
Spectra of Basic Pulsar Synthesis
Many time-varying parameters interact to produce the pulsar timbre, including
the pulsaret, the pulsaret envelope, the fundamental frequency, multiple formant frequencies, and the burst masking ratio. The spectrum of a single pulsar
stream is the convolution product of w and v, biased in frequency by fd and fp .

145

Varieties of Particle Synthesis

Figure 4.14 Pulsar rhythms. (Top) Pulse graph of rhythm showing rate of pulsar emission (vertical scale) plotted against time (horizontal scale). The left-hand scale measures
traditional note values, while the right-hand scale measures frequencies. (Bottom) Timedomain image of generated pulsar train corresponding to the plot above.

146

Chapter 4

Figure 4.15 Eect of the pulsaret envelope on the spectrum. The top row presents frequency-versus-time sonograms of an individual pulsar with a sinusoidal pulsaret, a fundamental frequency of 12 Hz, and a formant frequency of 500 Hz. The sonograms use
1024-point fast Fourier transform plots with a Von Hann window. They are plotted on a
linear frequency scale. From left to right, we see the sonogram produced by a rectangular envelope, an expodec envelope, and a Gaussian envelope. The lower row plots
the spectra of these pulsars on a dB scale.

Since w and v can be arbitrary waveforms, and fd and fp can vary continuously,
the range of spectra produced by PS is quite large.
When the formant frequency is set at a specic frequency energy spreads in
that region of the spectrum. Precisely how the energy spreads depends on w and
v. The pulsaret waveform w can be considered a template of spectrum shape
that repeats at the stipulated fundamental frequency fp and is time-scaled by
the duty cycle or formant frequency fd . If, for example, the ratio of the amplitudes of the rst ve harmonics of w is 5 : 4 : 3 : 2 : 1, this ratio prevails, independent of the specic values of p and d, when fp a fd .
The pulsaret envelope's contribution to the spectrum is signicant. Figure
4.15 shows the spectra of individual pulsars where the waveform w is a sinusoid,
and the pulsaret envelope v varies among three basic shapes. In the case of gure 4.15a, v is rectangular. Consequently, the formant spectrum takes the form
of a broad sinc sinx=x function in the frequency domain. The spectrum
shows strong peaks at factors of 1.5 fd , 2.5 fd , etc., and nulls at harmonics of fd .
This is characteristic of the sinc function. An exponential decay or expodec
envelope (such as in gure 4.15d) tends to smooth the peaks and valleys in

147

Varieties of Particle Synthesis

the spectrum (gure 4.15b). The bell-shaped Gaussian envelope compresses the
spectral energy, centering it around the formant frequency (gure 4.15c).
Thus by modifying the pulsaret envelope, one can alter the prole of the
pulsar spectrum. The appendix presents a mathematical analysis of the spectra
of simple pulsaret envelopes.
Advanced Pulsar Synthesis
Advanced pulsar synthesis builds upon basic pulsar synthesis by adding several
features that take it beyond the realm of vintage electronic sonorities. Three
methods are of particular importance:
1. Multiple pulsar generators sharing a common fundamental frequency but
with individual formant and spatial trajectories
2. Pulse-masking to shape the rhythm of the pulsar train
3. Convolution of pulsar trains with sampled sounds
Figure 4.16 outlines the schema of advanced pulsar synthesis. The following
sections explain the dierent parts of this schema.
Multiple Pulsar Generators
A pulsar generator has seven parameters:
1 Pulsar train duration
1 Pulsar train fundamental frequency envelope fp
1 Pulsaret formant frequency envelope fd
1 Pulsaret waveform w
1 Pulsaret envelope v
1 Pulsar train amplitude envelope a
1 Pulsar train spatial path s
The individual pulsar train is the simplest case. To synthesize a complex
sound with several resonance peaks, we can add several pulsar trains with the
same fundamental frequency but with dierent time-varying formant frequencies fd . One envelope controls their common fundamental frequency, while two
or more separate envelopes control their formant trajectories fd1 , fd2 , etc.

148

Chapter 4

149

Varieties of Particle Synthesis

A unique feature of pulsar synthesis is that each formant can follow its own
spatial path. This leads to complex spatial interplay within a single tone or
rhythmic phrase.
Pulsar Masking, Subharmonics, and Long Tonepulses
A pulsar generator emits a metronomic sequence of pulsars, where the rate of
emission can vary over time according to the fundamental frequency envelope
function fp . Pulsar masking breaks up the stream by introducing intermittancies
(regular or irregular) into the metronomic sequence. It deletes individual pulsarets, leaving an interval of silence in their place. This takes three forms: burst,
channel, and stochastic masking.
Burst masking (gure 4.17a) models the burst generators of the classic
electronic music studios. It produces a regular pattern of pulsarets that are
interrupted at regular intervals. The on/o pattern can be stipulated as
the burst ratio b : r, where b is the burst length in pulsaret periods and r is
the rest length in pulsaret periods. For example, a b : r ratio of 4 : 2 produces an alternating sequence of four pulsarets and two silent periods:
111100111100111100111100, etc. If the fundamental frequency is infrasonic, the eect is rhythmic.
When the fundamental is in the audio frequency range, burst masking
imposes an amplitude modulation eect on the timbre (gure 4.18), dividing
the fundamental frequency by a subharmonic factor b r. With the PulsarGenerator program (described later), we can alter the burst ratio in real time,
producing a gamut of subharmonic permutations.
When b r is large, the subharmonic crosses through the threshold separating tone and rhythm. The result is a series of alternating long tonepulses (at the
fundamental pitch) and silent intervals.
Channel masking (gure 4.17b) selectively masks pulsars in two channels, creating a dialog within a phrase by articulating each channel in turn.
Figure 4.17b shows two channels only, but we can generalize this scheme to N
channels.
Figure 4.16 Schema of pulsar synthesis. A pulsar generator with separate envelope
controls for fundamental frequency, formant frequency, amplitude, stochastic masking,
and spatial position. In advanced pulsar synthesis, several generators may be linked with
separate formant and spatial envelopes. A pulsar stream may be convolved with a sampled sound.

150

Chapter 4

Figure 4.17 Pulsar masking turns a regular train into an irregular train. Pulsars are
illustrated as quarter notes, and masked pulsars are indicated as quarter rests. (a) Burst
masking. The burst ratio here is 3 : 3. (b) Channel masking. (c) Stochastic masking
according to a probability table. When the probability is 1, there is no masking. When
the probability is 0, there are no pulsars. In the middle, the pulsar train is intermittent.
Notice the thinning out of the texture as the probability curve dips in the center.

151

Varieties of Particle Synthesis

Figure 4.18 Sonogram depicting the eect of burst masking in the audio frequency
range. The pulsaret is one cycle of a sinusoid, and the pulsaret envelope is rectangular.
The b : r ratio is 2 : 1. The fundamental frequency is 100 Hz and the formant frequency is
400 Hz. Notice the subharmonics at 133 Hz and 266 Hz caused by the extended periodicity of the pulse masking interval (400 Hz/3).

Stochastic masking introduces random intermittancy into the regular stream

of pulsars. We have implemented stochastic masking as a weighted probability
that a pulsar will be emitted at a particular point in a pulsar train. The probability is expressed as an envelope over the duration of the pulsar train. When
the value of the envelope is 1, a pulsar is emitted. If the value is less than 1, it
has less possibility. A value of 0 results in no pulsar emissions. Values between
0.9 and 0.8 produces an interesting analog-like intermittancy, as if there were
an erratic contact in the synthesis circuit (gure 4.17c).

152

Chapter 4

Convolution of Pulsars with Samples

Pulsar synthesis can be harnessed as a method of sound transformation through
convolution. Convolution is fundamental to the physics of waves (Rabiner
and Gold 1975). It ``crosses'' two signals, creating a new signal that combines
the time structures and spectra of both inputs. Many transformations emerge
from convolution, including exotic lters, spatializers, models of excitation/
resonance, and a gamut of temporal transformations (echoes, reverberation,
attack-smoothing, rhythm-mapping). Pure convolution, however, has no control parameters, that is, the type of eect achieved depends entirely on the
nature of the input signals. See Roads (1992b, 1993a, 1997) for applications of
convolution in musical sound transformation.
Sophisticated transformations involving rhythm- and spatial-mapping can be
achieved through convolution. It is well known that any series of impulses
convolved with a brief sound maps that sound into the time pattern of the
impulses. These impulses can be emitted by a pulsar generator. If the pulsar
train frequency is in the infrasonic range, then each pulsar is replaced by a copy
of the sampled sound object, creating a rhythmic pattern. The convolution of a
rhythmic pattern with a sound object causes each impulse to be replaced by a
ltered copy of the sound object. Each instance of the sampled object is projected in space according to the spatial location of a specic pulsar's position.
In convolution, each pulsar represents the impulse response of a lter. Thus
timbral variations can derive from two factors: (1) ltering eects imposed by
the time-varying pulsar train, and (2) overlapping eects caused by convolution
with pulsar trains whose fundamental period is shorter than the duration of the
sampled sound.
Figure 4.19 shows the temporal and ltering eects of convolution in the
form of sonograms. The input signal (b) is the Italian word qui (pronounced
``kwee''). It convolves with the pulsar train (a) with a variable infrasonic
fundamental frequency and a variable audio formant frequency. The resulting
convolution (c) combines the time structure and the spectra of the two signals.
The composer can stockpile a database of sampled sound objects for crossing
with trains selected from the pulsar database. If the goal of the synthesis is to
retain the time structure of the pulsar train (e.g., to maintain a specic rhythm),
the sampled sound objects should be of short duration (less than the fundamental period of the pulsar train) and have a sharp attack (a rise time of less
than 100 ms). These constraints minimize the time-smearing eects of convolution (Roads 1992b, 1993a, 1997). A good starting point for a sound database

153

Varieties of Particle Synthesis

Figure 4.19 Eect of convolution with pulsar train. (a) Infrasonic pulsar train with a
variable fundamental and formant frequency. (b) Sampled sound, the Italian word qui
(pronounced ``kwee''). (c) Convolution of (a) and (b).

154

Chapter 4

is a collection of percussion samples. If one seeks a smoother and more continuous texture the constraints can be relaxed. Samples with long durations
superimpose multiple copies of the sampled object, creating a rippling sound
stream. Samples with slow attacks blur the onset of each sample copy, smearing
the stream into a continuum. Thus by controlling the attack shape of the sample one can aect the sonic texture.
Implementations of Pulsar Synthesis
My original implementation of PS dates to 1991, using James McCartney's
Synth-O-Matic, a programmable sound synthesis environment for Apple
Macintosh computers (McCartney 1990, 1994). In 1996, Mr. McCartney replaced Synth-O-Matic with SuperCollider 1an object-oriented programming
language with a Power Macintosh runtime system (McCartney 1996). Using
SuperCollider 1, Stephen T. Pope and I created a new implementation of basic
PS in 1997.
With the improved SuperCollider 2 (McCartney 1998), Alberto de Campo
and I developed a new realization of pulsar synthesis, presented in a 1999
summer course at the Center for New Music and Audio Technology, University of California, Berkeley. Further renement of this prototype has led to the
PulsarGenerator application, distributed by CREATE. Figure 4.20 shows the
graphical interface of PulsarGenerator, version 1. Notice the control envelopes
for the synthesis variables. Users can design these envelopes in advance of
synthesis, or manipulate them in real time as the instrument plays. We have
implemented a scheme for saving and loading these envelopes in groups called
settings. The program lets one crossfade at a variable rate between settings,
which takes performance with PulsarGenerator to another level of musical
complexity.
In wave-oriented synthesis techniques, an algorithm loops through a wavetable and varies the signal according to relatively slowly-updated control
functions. Thus the eciency of synthesis corresponds to the number of simultaneous unit generators (oscillators, lters, etc.). In contrast, particle synthesis is more demanding, since the synthesis algorithm must also handle the
task of scheduling possibly thousands of events per second, each of which may
be unique. The eciency of pulsar synthesis is therefore related to the rate of
particle emission. At infrasonic rates (< 20 pulsars per second), the PulsarGenerator application uses less than 3.6% of the processor on a single-processor
Apple G4 running at a 500 MHz clock speed. At high audio rates (such as
a three-formant instrument emitting six thousand pulsars per second, corre-

155

Varieties of Particle Synthesis

Figure 4.20 Control panel of the PulsarGenerator application by Alberto de Campo

and Curtis Roads. Copyright Alberto de Campo, Curtis Roads, and the Regents of the
University of California 2000.

sponding to the fundamental frequency of 2 kHz), the application requires approximately 45% of the processor. It is a testimony to SuperCollider 2 that the
entire implementation, including the graphical interface, needed less than one
thousand and ve hundred lines of code and comments. Our code builds the
interface, denes the synthesis algorithm, schedules the pulsars, and handles le
input and output. McCartney's SCPlay, an ecient real-time sound engine,
calculates the samples.
Composing with Pulsars
To interact with PulsarGenerator in real time is to experiment directly with
sonic ideas. While experimenting, a composer can save settings and plan how
these will be used within a composition. The PulsarGenerator program can also

156

Chapter 4

record the sounds produced in a real-time session. The composer can then edit
the session or convolve and mix it with other material.
A nal stage of pulsar composition is to merge multiple trains to form a
composite texture. This is a question of montage, and is best handled by editing
and mixing software designed for this purpose. Each layer of the texture may
have its own rhythmic pattern, formant frequency envelope, choice of convolved objects, and spatial path. Working on a variety of time scales, a composer can apply signal processing transformations such as mixing with other
sounds, ltering, modulations, and reverberation to individual pulsars, pulsar
trains, and pulsar textures.
Musical Applications of Pulsar Synthesis
I developed pulsar synthesis while realizing Clang-Tint (Roads 1993b), an electronic music composition commissioned by the Japanese Ministry of Culture
(Bunka-cho) and the Kunitachi College of Music, Tokyo. The second movement of this work, entitled Organic, focuses on expressive phrasing. It combines
bursts of insect noise and animal and bird calls with electronic pulse-tones. The
electronic sound palette utilizes pulsar synthesis in many forms: pulsating blips,
elongated formant tones, and clouds of asynchronous pulsars. For the latter, I
rst generated multiple infrasonic pulsar trains, each one beating at a dierent
frequency in the range of 6 to 18 Hz. I then mixed these together to obtain the
asynchronous pulsar cloud.
The raw material of my electronic music composition Half-life, composed in
1998 and 1999, is a one-minute pulsar train that is wildly varied. Most of the
sounds in the rest of the work are derived from this source. Half-life extends
the pulsar material through processes of granulation, microltration, granular
pitch-shifting, recirculating feedback echo, individual pulsar amplitude shaping,
and selective reverberation. Tenth vortex (2000) and Eleventh vortex (2001)
continue in this direction.
Assessment of Pulsar Synthesis
Music unfolds on multiple time scales, from high-level macrostructure down to
a myriad of individual sound objects or notes. Below this level is another hierarchy of time scales. Here are the microsonic particles such as the classical
rectangular impulses, grains, wavelets, and pulsars (Roads 1999). Musicians
proved the eectiveness of analog impulse generation decades ago. In com-

157

Varieties of Particle Synthesis

parison, digital pulsar synthesis oers a exible choice of waveforms and envelopes, increased precision, and graphical programmable control.
Unlike wave-oriented synthesis techniques, the notion of rhythm is built into
techniques based on particles. Rhythm, pitch, and timbre are all interrelated
but can be separately controlled. Pulsar synthesis oers a seamless link between
the time scales of individual particle rhythms, periodic pitches, and the meso or
phrase level of composition. A novel feature of this technique is the generation
of multiple independent formant trajectories, each following its own spatial path.
As we have shown, basic pulsar technique can create a broad family of
musical structures: singular impulses, rhythmic sequences, continuous tones,
time-varying phrases, and beating textures. Pulsar microevents produce rhythmic sequences or, when the density of events is suciently high, sustained tones,
allowing composition to pass directly from microstructure to mesostructure.

Graphic and Sonographic Synthesis of Microsound

Graphic and sonographic sound synthesis employ visual means to translate
images into sound. Here we look at two approaches to interpreting graphical
images as sound waveforms: time-domain or graphic, and frequency-domain or
sonographic.
Graphic Synthesis in the Time-Domain and Frequency-Domain
Electro-optical synthesizers read waveforms inscribed on rotating tone wheels
as in a recent implementation of photosonic synthesis using a rotating optical
disc with its inscribed time-domain waveform (Duden and Arb 1990; Arb,
Duden, and Sanchez 1996). In this approach to graphic synthesis, a short
waveform repeats many times per second to form a tone.
Another approach is to draw a waveform over an extended period of time. A
sound editor presents an image of sound waveforms in a time domain, some
programs providing an onscreen pencil for drawing waveforms or smoothing
over discontinuities. The time-domain approach to graphic synthesis can be
eective on the micro time scale, but requires a great deal of labor. In the era
of optical lm soundtracks, a few dedicated souls made short pieces using a
similar technique. A prime example is Norman McLaren's 1948 Dots Points, a
2-minute 47-second lm in which both sound and visuals are drawn directly on
the lm with pen and ink. (See also the transient drawing technique presented

158

Chapter 4

later in this chapter, which creates individual particles by sculpting waveforms

in a sound editor.)
Time-domain graphic synthesis lets one design the microstructure of each
sound event in detail. However, this approach does not scale well to higher time
spans. Acoustical microstructure is delicate. The ear is extremely sensitive to
even minor alterations in transient morphology and envelope shape, such as the
slope or duration of an attack. Even a slight amplitude discontinuity may cause
an audible click or thump. At the same time, the ear tends to ignore permutations in the phase of partials that cause dramatic changes in the shape of the
waveform.
Waveform drawing loses usefulness as one zooms out from a waveform display to take in a larger time view. One gains a global view of the sound object
envelope, but can no longer see the inner structure of the sound (waveshape,
number of iterations, microvariations, etc.), which largely determine its identity.
Another approach to graphic synthesis, which is more powerful on larger
time scales, is based on a sonogram frequency-domain representation. To produce a sonogram, a sound is passed through a bank of lters and the output
energy from each lter is traced as a function of frequency and time. The
sonogram projects frequency on the vertical axis and time on the horizontal
axis. The amplitude appears as darkness on a white background, where a
darker trace means more energy.
Any visual representation that captures the basic idea of a frequency-versustime plane is a sonographical notation. Now available are a number of systems,
where graphical traces which one can inscribe on a frequency-time grid can be
converted directly into sound. In eect one can paint the sound, erase it, or
touch it up with all the exibility of a software paint program. Indeed, several
systems let one import images that have been prepared with visual software.
The next sections survey three sonographic systems, with emphasis on their
microsonic capabilities.
Micro-Arc Synthesis with the UPIC
The UPIC (Unite Polyagogique Informatique de CEMAMu) is a sound synthesis system conceived by Iannis Xenakis and built by engineers at the Centre
d'Etudes de Mathematique et Automatique Musicales (CEMAMu) in Paris
(Xenakis 1992). The UPIC system oers both graphical (time-domain) and
sonographical (frequency-domain) interfaces for sound composition. In an initial version, dating from 1977, the user interacted by way of a large graphics

159

Varieties of Particle Synthesis

tablet, mounted vertically like a painter's easel. The rst composition realized
with the UPIC was Xenakis's Mycenae-Alpha (1980). A major breakthrough
for the system was the development of a real-time version, based on a 64oscillator synthesis engine (Raczinski and Marino 1988). By 1991, engineers
had coupled this engine to a personal computer running the Windows operating
system, permitting a sophisticated graphical interface (Marino, Raczinski, and
Serra 1990; Raczinski, Marino, and Serra 1991). The program now runs standalone, with no additional hardware.
At the level of sound microstructure, waveforms and event envelopes can
be drawn directly onto the tablet and displayed onscreen. At a higher level of
organization, composers can draw the sonographical frequency/time structure
of a score page. Linescalled arcsappear on the display screen as one draws
with the mouse. Individual arcs can then be moved, stretched or shrunk, cut,
copied, or pasted. The arcs on the page can also represent sampled sounds.
In the 1991 version of the UPIC, a page can have sixty-four simultaneous
arcs, with four thousand arcs per page. Most importantly, the duration of each
page can last from 6 ms to more than two hours. This temporal exibility lets
the user zoom in to the micro time scale. When a page lasts only a second, say,
any arcs written onto it will be microsounds. These micro-arcs can also be cut,
copied, and pasted, as well as stretched or compressed in time and frequency.
Moreover, the rate at and direction in which the score is read can be controlled
in real time with a mouse. This allows discontinuous jumps from one region of
the score to another, for example. The sequence of control motions as it plays
a score can be recorded and later the same performance can be replayed or
edited.
The UPIC system is an especially pliable musical tool since it integrates
many levels of composition within a common user interface. Graphic functions
created onscreen can function equally as envelopes, waveforms, pitch-time scores,
tempo curves, or performance trajectories. This uniform treatment of composition data at every level should be extended to more computer music systems.
Synthesis of Microsound in Phonogramme
Vincent Lebros developed Phonogramme in 1993 at the Universite de Paris
VIII (Lesbros 1995, 1996). Phonogramme oers an approach to graphical synthesis with some similarities to the UPIC, but oering a number of extensions.
First, the program can generate sound directly from a MacOS computer, or it
can generate MIDI data to be sent to a bank of synthesizers. Second, the pro-

160

Chapter 4

gram oers a harmonic pen, which draws a number of harmonically related

partials above the arc inscribed by the user. Users can adjust the number and
strength of these partials before drawing. The harmonic pen solves a common
problem of sonographic synthesis: a single arc is not a complete description of a
sound object. As a sonogram of a concrete sound clearly indicates, it takes
many traces on the sonographic plane to create a sound object with a fundamental frequency, formant structure, harmonic/inharmonic partials, and inner
variations. Finally, unlike the 1991 UPIC system, Phonogramme can analyze a
sound and generate an editable sonographic representation.
Phonogramme's score page can be interpreted on any time scale chosen by
the user. A line or arc inscribed can be interpreted as a microsonic particle. It is
also possible to turn the drawing pencil into a kind of dot spray where a fast
gesture made with the mouse leaves behind not a continuous line but a series of
microsonic dots (gure 4.21).
Synthesis of Microsound in MetaSynth
The MetaSynth program by (Wenger and Spiegel 1999) is a recent implementation of sonographical synthesis. It operates on a MacOS computer with realtime playback (or delayed playback for complex scores). Although it does not
incorporate every feature of predecessors such as UPIC and Phonogramme, it
does oer unique features that take graphical synthesis to a new level of power
and expressivity.
Among the drawing implements included with MetaSynth are some tools for
microsound synthesis, including a granular spray brush (gure 4.22). The spray
brush determines the time and frequency span of the grains. The grain envelope
is xed as a triangle. In another mode, it can play back a line drawing with
a granular instrument. The waveform of the grains is variable on a per-page
basis, from a synthetic waveform to a sampled waveform.
MetaSynth is a major contribution to sonographic synthesis in that it greatly
expands the palette of brushes and other implements to paint sound color. It
also permits nongraphical granulation of sound les in its Eects window. (See
chapter 5 for more on granulation of sound les.)
Assessment of Graphic and Sonographic Synthesis of Microsound
Graphic and sonographic synthesis share a common interface: drawing. What
their drawings represent are quite dierent, however. Graphic synthesis depicts

161

Varieties of Particle Synthesis

Figure 4.21 Phonogramme scores. Both scores are just over 4.6 seconds in length.
(a) Fast horizontal gestures leave behind a stream of micro-arcs. Notice the four harmonics superimposed over the original low-frequency gestures by the harmonic pencil.
(b) Slow hand movements create continuous tones.

162

Chapter 4

Figure 4.22 Sonographical synthesis of particles in MetaSynth. (a) The particles were
drawn by hand with a spray brush tool. This particle score can be played back with
many dierent waveforms, including sampled sounds. (b) Time-domain view of the
waveform.

time-domain envelopes and waveforms. To draw pitch curves and envelope

shapes with graphic tools is a supple and ecient procedure. It is easy to create
shapes and phrases that would be dicult to achieve by other means.
To draw time-domain waveforms, however, is problematic. It is dicult to
infer a sound directly from a view of its waveform. This is further complicated
by the fact that the same sound can be represented graphically by innumerable
ways whose only dierences are in the phase relationships among the frequency
components. Furthermore, any waveform repeated without variation takes on a
static quality. So, as in other approaches to synthesis, waveform generation in
graphic synthesis systems has shifted from individual xed waveforms to
evolving sources such as sampled sounds or groups of time-varying waveforms.
Sonographic sound synthesis is a direct and intuitive approach to sound
sculpture. Interaction with sonographical synthesis can be either precise or imprecise, depending on how the user treats the process. A composer who plans

163

Varieties of Particle Synthesis

each line and its mapping into sound can obtain precise results. It is just as
worthwhile to treat the medium as a sketch pad, where initial drawings are later
rened into a nished design.
Unlike traditional notation, which requires serious study for a long period of
time, a child can learn the relationship between drawn gestures and sound in
minutes. This initial simplicity hides a deeper complexity, however. As with any
technique, the best results demand a long period of study and experimentation.
As mentioned earlier, a single arc is not a description of a complete sound
object. An arc is only one component of a complex time-varying sound. Such
sounds require many arcs. We can see this complex nature in the sonograms of
relatively simple instrumental tones, such as a cello. Noiser timbres, such as a
cymbal or gong, display enormous complexity. They seem to be composed of
globules of energy that sometimes connect and at other times break apart.
As such representations become more commonplace, a long process of codicationfrom complex sonographic patterns into abstract iconic notation
seems to be inevitable. The Acousmagraphe system, developed at the Groupe
de Recherches Musicale (Paris), points in this direction (Desantos 1997).

Particle-Based Formant Synthesis

A formant is a peak of energy in a spectrum, which can include both harmonic
and inharmonic partials as well as noise. Formant peaks are a characteristic of
the spoken vowel sounds and the tone color of many musical instruments.
Within the range of 0 to 5000 Hz, the vocal tract has ve formant regions.
Formant regions act as a kind of ``spectral signature'' or ``timbral cue'' to the
source of many sounds. The formants of a voice or an instrument are not xed,
however, they drift according to the frequency of the fundamental (Luce 1963).
Furthermore, formants are only one clue the ear uses to identify the source of a
tone.
Fully understanding the nature of formants in human speech has long been a
goal of scientic research. Ingenious methods for synthesizing the formants of
vowel-like tones have been developed (Tyndall 1875; Miller 1916), and it is not
surprising that this research has served as a wellspring of ideas for musical formant synthesis.
The rest of this section presents concise descriptions of three particle synthesis techniques: formant wave-function or FOF synthesis, Vosim, and windowfunction (WF) synthesis. FOF and Vosim evolved from attempts to simulate

164

Chapter 4

human speech, whereas WF was developed to emulate the formants of traditional musical instruments. For more detailed descriptions, see the references
and Roads (1996).
FOF Synthesis
Formant wave-function synthesis ( fonction d'onde formantique or FOF) generates a stream of grains, each separated by a quantum of time, corresponding
to the period of the fundamental frequency. So a single note produced by this
technique contains hundreds of FOF grains. The denitive FOF grain is a sine
wave with either a steep or smooth attack and a quasi-exponential decay.
The envelope of a FOF grain is local, that of the entire note is global. The
local envelope of a FOF grain is dened as follows. For the attack portion of
the FOF grain, 0 a t a tex, the envelope is:
envt 1=2  1 cospt =tex  expattent
For the decay portion, t b tex, the envelope is:
envt expattent
p is the initial phase of the FOF signal, tex is the attack time of the local envelope, and atten is the decay time (D'Allessandro and Rodet 1989). The eect
is that of a damped sinusoidal burst, each FOF grain lasting just a few milliseconds. The convolution of the brief FOF envelope with the sinusoid contributes audible sidebands around the sine wave, creating a formant spectrum. The
spectrum of the damped sine generator is equivalent to the frequency response
curve of one of the bandpass lters and the result of summing several FOF
generators is a spectrum with several formant peaks.
Each FOF generator is controlled by many parameters. Among these are the
formant parameters p1 through p4:
p1 is the center frequency of the formant.
p2 is the formant bandwidth, dened as the width between the points that are
6 dB from the peak of the formant.
p3 is the peak amplitude of the formant.
p4 is the width of the formant skirt. The formant skirt is the lower part of the
formant peak, about 40 dB below the peak, akin to the foothills of a mountain.
The skirt parameter is independent of the formant bandwidth, which species
the breadth at the peak of the mountain.

165

Varieties of Particle Synthesis

The inherent connection between time-domain and frequency-domain operations is exemplied in the way FOF parameters are specied. Two of the main
formant parameters are specied in the time domain as properties of the envelope of the FOF grain. First, the duration of the FOF attack controls parameter p4, the width of the formant skirt, that is, as the duration of the attack
lengthens, the skirtwidth narrows. Second, the duration of the FOF decay
determines p2, the formant bandwidth. Hence a long decay length translates
into a sharp resonance peak, while a short decay widens the bandwidth of the
signal.
The basic sound production model embedded in FOF synthesis is the voice.
However, users can tune many parameters to move beyond vocal synthesis
toward synthetic eects and emulations of instruments (Bennett and Rodet
1989).
Typical applications of FOF synthesis congure several FOF generators in
parallel. Some implementations are very complicated, with over 60 parameters
to be specied for each sound event. The CHANT program, developed in the
1980s, was proposed as a response to this complexity, providing a collection
of rules for controlling multiple FOF streams in parallel (Rodet, Potard, and
Barriere 1984).
Vosim
Like FOF synthesis, the Vosim technique generates a series of short-duration
particles in order to produce a formant eect. Vosim synthesis was developed
by Werner Kaegi and Stan Tempelaars at the Institute of Sonology in Utrecht
during the early 1970s (Kaegi 1973, 1974; Tempelaars 1976, 1977, 1996). Vosim
generates a series of tonebursts, producing a strong formant component. Like
FOF, it was originally designed for vowel sounds, and later extended to model
vocal fricativesconsonants such as [sh]and quasi-instrumental tones (Kaegi
and Tempelaars 1978).
The Vosim waveform approximates the signal generated by the human voice
in the form of a series of pulsetrains, where each pulse is the square of a sine
function. The parameter A sets the amplitude of the highest pulse. Each of the
pulsetrains contains N sin 2 pulses in series decreasing in amplitude by a decay
factor b. The width (duration) of each pulse T determines the position of the
formant spectrum. A variable-length delay M follows each pulse train, which
contributes to the pulsetrain's overall period, and thus determines the fundamental frequency period. The period is N  T M, so for seven pulses of

166

Chapter 4
Table 4.3 VOSIM parameters
Name

Description

T
dT
M
dM
D
A
dA
b
N
S
NM
NP

Pulsewidth
Increment or decrement of T
Delay following a series of pulses
Increment or decrement of M
Maximum deviation of M
Amplitude of the rst pulse
Increment or decrement of A
Attentuation constant for the series of pulses
Number of pulses per period
Type of modulation (sine or random)
Modulation rate
Number of periods

200 msec duration and a delay equal to 900 msec, the total period is 3 ms and
the fundamental frequency is 333.33 Hz. The formant centers at 5000 Hz.
Two strong percepts emerge from the typical Vosim signal: a fundamental
corresponding to the repetition frequency of the entire signal, and a formant
peak in the spectrum corresponding to the pulsewidth of the sin 2 pulses. A
Vosim oscillator produces one formant. In order to create a sound with several
formants, it is necessary to mix the outputs of several Vosim oscillators.
Table 4.3 lists the set of parameters that control the Vosim oscillator. T, M,
N, A, and b are the primary parameters. By modulating the delay period M,
one can produce vibrato, frequency modulation, and noise sounds. Kaegi and
Tempelaars introduced three additional variables: S, D, and NM, corresponding respectively to the type of modulation (sine or random), the maximum frequency deviation, and the modulation rate. They wanted also to be able to
provide for ``transitional'' sounds, which led to the introduction of the variables
NP, dT, dM, and dA. These are the positive and negative increments of T, M,
and A, respectively, within the number of periods NP.
By changing the value of the pulsewidth T, the formant changes in time. This
is formant shifting, a dierent eect than the progressive spectral enrichment
which occurs in, for example, frequency modulation synthesis. The Vosim signal is not bandlimited, but spectral components are greater than 60 dB down at
six times the fundamental frequency (Tempelaars 1976).

167

Varieties of Particle Synthesis

Window Function Synthesis

Window function (WF) synthesis generates pulsestreams that result in formants
of purely harmonic content (Bass and Goeddel 1981). It begins with the creation of a broadband harmonic signal. Then a weighting stage emphasizes or
attenuates dierent harmonics in this signal to create time-varying formant
regions which can emulate the harmonic spectra of traditional instruments.
The building block of WF synthesis is a particle known as a window function
pulse. A typical window function is a smooth pulse with a bell-shaped envelope.
Window spectra exhibit a characteristic center lobe and side lobes. The center
lobe is typically much higher in amplitude than the side lobes, meaning that the
signal is bandlimited. In the Blackman-Harris window function chosen by Bass
and Goeddel, the frequencies in the side lobes are down at least 60 dB.
Bass and Goeddel created a broadband signal by linking a periodic series of
WF pulses separated by a period of zero amplitude. For dierent fundamental
frequencies, the duration of the WF pulse stays the same; only the interpulse
silence varies. In this use of a pulse followed by a period of deadtime, the WF
technique is similar to pulsar synthesis, Vosim, and the FOF method explained
earlier. In other respects, however, the techniques are not similar.
In WF synthesis, the number of harmonics increases as the fundamental frequency decreases. This is because the higher harmonics fall outside the center
lobe of the WF pulse's spectrum. Thus, low tones are timbrally rich, while high
tones are less so. This is characteristic of some traditional instruments such as
pipe organs and pianos, which Bass and Goeddel wanted to simulate. Note
that not all instrumentsfor example, the harpsichordexhibit this behavior.
Furthermore, some instruments do not have a purely harmonic spectrum and
so make poor models for the WF technique.
To create formant regions in the spectrum requires a further processing called
slot weighting. A time slot is dened as the duration of a single WF pulse plus
a portion of its deadtime. The timbre of the sound can be altered by weighting the slots (i.e., multiplying a slot by a value) with a periodic sequence of
N slot weights. This feeds a stream of WF pulses to a multiplier as an input
signal, along with a periodic stream of slot weights. The multiplier computes
the product of each input pulse with a specic weight, making an output stream
of WF pulses at dierent amplitudes. The spectrum of such a stream exhibits
peaks and valleys at various frequencies. For time-varying timbres, each slot
weight can be specied as a time-varying function.

168

Chapter 4

WF synthesis requires an amplitude compensation scheme, because low frequencies contain few pulses and much zero-amplitude deadtime or silence,
while high frequencies contain many pulses and almost no deadtime. A quasilinear scaling function adjusts the amplitude as an inverse function of frequency. That is, low tones are emphasized and high tones are attentuated for
equal balance throughout the frequency range.
Assessment of Particle-Based Formant Synthesis Techniques
Particle-based formant synthesis models a class of natural mechanisms which
resonate when excited, and are quickly damped by physical forces. A typical
example would be a stroke on a woodblocka sound which cuts o almost
immediately. The result is a grain-like ``pop.'' Another example is the glottal
pulse, which the vocal tract lters. Continuous tones string together a series of
such particles.
FOF synthesis has been available within the widely distributed Csound language for some time (Boulanger 2000). The Common Lisp Music language
(Schottstaedt 2000) includes a wavetrain object able to realize both FOF and
Vosim synthesis. Window Function synthesis technique was experimental, and
since its original realization, has not continued.
Vocal-like tones can be simulated by mimicking the fast impulses that continuously excite resonance in the vocal tract. Realistic simulation from a particle technique, however, requires an enormous investment of time. In the 1980s,
vocal synthesis using the FOF technique performed the ``Queen of the Night''
aria from Mozart's Magic Flute. The realization of this 30-second fragment
took months of eort.
In the simulation of vocal and instrumental tone, the particle representation
should be invisible. If we divorce these techniques from their original uses, we
can see that particle-based formant synthesis remains a rich resource for synthetic timbres, both at the infrasonic frequency levelwhere it produces a wide
variety of rhythmic popsand also at the audio frequency level, where it generates expressive resonant tones.

Synthesis by Transient Drawing

Synthesis by transient drawing is a direct method. It takes two forms: synthetic
and transformational. Synthetic transient drawing requires a graphical sound

169

Varieties of Particle Synthesis

Figure 4.23 Transient wave writing in a sound editor. Notice the hand-drawn transients
interspersed with narrow computer-generated impulses.

editor which provides a waveform pencil that can be used to manually inscribe
the waveform of a transient. These transients can be interspersed with waveforms produced by other means (gure 4.23).
Transformational transient drawing also uses a sound editor, after beginning
with an existing low-level audio signal, such as background noise or the tail of a
reverberant envelope. A brief extract of this nonstationary signal is selected and
rescaled to a much higher amplitude, creating a transient particle (gure 4.24).
This particle can be reshaped using the implements of the editor, such as narrowband ltering, envelope reshaping, phase inversion, and spatialisation. The
following frequency bands are especially important in ltering the transient
particles.
1. Direct current (DC) cutremoval of frequencies below 20 Hz
2. Deep bass cut or boost80 Hz
3. Mid-bass boominess cut200 Hz
4. Low-mid boost500700 Hz resonances
5. Mid harshness cut12 kHz

170

Chapter 4

Figure 4.24 Manipulations of a transient wave extracted from an ambient texture.

(a) Ambient noise texture. The darkened part is a 120-ms transient that we extract as a
transient wave. (b) Detail of the extracted transient. (c) Reversal of (b). (d) Lowpass ltered (smoothed) version of (b). (e) Re-enveloped and pitch-shifted copy of (b), now
lasting less than 50 ms. (f ) Ten-fold replication (cloning) of (e) to form a decaying sound
object. (g) Copy of (b) that has been passed through a bank of narrow lters to form a
pitched tone. The attack was edited to create a sharper onset.

6. Crackle boostnotches at 3 kHz and 6 kHz

7. High harshness removallowpass at 5 kHz
8. Brilliance boost12 kHz
Filters 2, 4, 6, and 8 accentuate the particle, while lters 1, 3, 5, and 7 attenuate
it in the harsh frequency zones. Successive micro operations can radically
transform the original signal. For example, a sharp narrow lter applied to a
brief noise particle transforms it into a resonating pure tone (gure 4.24g). If we
apply advanced techniques such as time-stretching, there comes a point at which
we can transform any transient into a myriad of sounds on any time scale.

171

Varieties of Particle Synthesis

Assessment of Transient Drawing

Transient drawing takes compositional decision-making down to a microscopic
level, starting from almost nothing and obliging the composer to handcraft
each individual particle in a phrase. The extreme dierentiation resulting from
this approach leads to a high probability of singularities or unique sound
events. In this sense, transient drawing maximizes the information content of
the musical signal.
Both the synthetic and the transformational approaches to transient drawing
are labor-intensive, relying on manual editing. In eect, the composer works in
a manner similar to a Pointillist painter such as Georges Seurat, building an
image from thousands of tiny brush strokes. In Seurat's later paintings (for
example, La Luzerne, Saint-Denis) the particles can be scattered within dense
masses. When one of these particles appears isolated by silence, its singularity
conveys a striking impact.
As we saw in the section on graphic synthesis, waveform drawing makes
sense only on a micro time scale. It does not translate well to higher time scales.
This is because zooming out on a waveform display provides a global view of
the sound object envelope at the expense of the inner microstructure (waveform
shape, number of iterations, etc.).
Transient drawing has become an integral part of my compositional practice.
Many examples of it appear in Clang-Tint, especially in the movement Organic.
I also made extensive use of the technique in the rst movement of Half-life, in
which a number of singularities appear. Perhaps the most practical use of
transient drawing is for fabricating such singularities to be scattered over sparse
textures. In order to build continuous tones from a single transient, I used
transient drawing in conjunction with particle cloning (described next).

Particle Cloning Synthesis

In particle cloning synthesis, a brief tone pip, or a longer continuous or uttering tone, is made by repeatedly cloning a single sound particle. This particle
may derive from any synthesis technique, including transient drawing. Figures
4.24f and 4.25 illustrate particle cloning. Here is a recipe for particle cloning:
1. Generate a single particle by any means.
2. Sculpt the particle to the desired shape and duration in a waveform editor.

172

Chapter 4

Figure 4.25 Particle cloning synthesis. (a) Solo particle, lasting 35 ms, extracted from
an acoustic sample of a drum sound. (b) A 200 ms sound object formed by cloning (a) 50
times, pitch-shifting it up two octaves, creating another channel slightly delayed, and
applying an exponentially decaying amplitude curve.

3. Clone the particle and repeat it over a specied duration to form a tone pip.
The number of particles cloned corresponds to the total duration divided by
the duration of the particle. The resulting pitch depends on the period of the
particle.
4. Shape the amplitude envelope of the resulting tone pip.
5. Pitch-shift the tone pip to the desired pitch, with or without time correction.
6. Apply one or more bandpass lters. The important bands to control are the
low (50150 Hz), low-mid (150300 Hz, narrowband), mid (500900 Hz),
mid-high (34 kHz, wide bandwidth), and high (912 kHz, wide bandwidth)
ranges.
In stage 3, we used a Replicate function in a sound editor to automate the
process. Replicate lls an arbitrary duration with the contents of the clipboard.
Thus it is possible to copy a particle, select a 3-second region, and ll the region
with a sequence of cloned particles. Obviously, the frequency of the tone pip
depends on the period of the particle, and so a 10-ms particle produces a tone

173

Varieties of Particle Synthesis

pip at a pitch of 100 Hz. By selecting some of the silent samples around a particle, one can shift the fundamental frequency downward. The resulting tone
can then be transposed to any desired pitch. If the period of the particle is
greater than about 50 ms, the texture is no longer continuous, but utters.
Stages 4, 5, and 6 foster heterogeneity among tones cloned from the same
particle. Each tone can have a unique duration, amplitude envelope, pitch, and
spectrum weighting.
Assessment of Particle Cloning Synthesis
The construction of tones by particle cloning elevates the time scale of a particle
from the micro level to the sound object level, and so fosters the emergence of
singularities on the sound object time scale. As we have implemented it, it is a
manual technique, carried out with a sound editor. I developed the particle
cloning method in the course of composing Half-life (199899), in which the
initial particles derive from pulsar synthesis or transient drawing. These particles stand out prominently in part 1 of the piece, in the melodic section at time
2:042:11.

Physical Models of Particles

Computer sound synthesis rests on the premise that a software program can
model a sound-generating process. Physical modeling (PhM) synthesis starts
from a mathematical description of acoustic sound production (Roads 1996).
That is, the equations of PhM describe the mechanical and acoustic behavior
of an instrument as it is played. The goals of physical modeling synthesis are
both scientic and artistic. PhM investigates the extent to which equations
and algorithms can simulate the sound-producing mechanisms of existing instruments. In this sense, a physical model embodies the Newtonian ideal of a
mathematical model of a complicated mechanico-acoustic process. PhM can
also create sounds of fancifully imaginary instruments, otherwise impossible to
realize.
PhM synthesis can be traced back to the mathematical descriptions of John
William Strutt (Lord Rayleigh). His insightful volume The Theory of Sound
(1894/1945) laid the foundation for a century of research. For many years,
eorts in physical modeling focused on synthesis of continuous tones. Only

174

Chapter 4

Figure 4.26 Perry Cook's model of a maraca, coded using the Synthesis Toolkit in the
C language. Notice the declaration at the top indicating the number of beans in the
shaker. A statement in a score le (not shown) triggers the maraca model.

more recently has attention turned to synthesis using physical models of the
particulated sounds of certain percussion instruments and environmental
microsounds. Perry Cook's Physically Informed Stochastic Event Modeling
(PhISEM) exemplies this approach. PhISEM is a suite of programs to simulate the sounds of shaken and scraped percussion such as maracas (gure 4.26),
sekere, cabasa, bamboo windchime, tambourine, sleighbells, and guiro (Cook
1996, 1997). Cook also developed a model to simulate the sound of water drops
based on the same principles, and suggested that this technique could also synthesize the sound of feet crunching on gravel, or ice cubes in a shaken glass.
(See also Keller and Truax (1998) and the discussion of physical models for
granular synthesis in chapter 3).
The common thread among instruments modeled by PhISEM is that sound
results from discrete microevents. At the core of PhISEM are particle models.

175

Varieties of Particle Synthesis

Basic Newtonian equations governing the motion and collision of point masses
produce the sounds. For shaken percussion instruments, the algorithm assumes
multiple individual sound sources, for example, that with maracas each of them
contains many beans. It calculates the probability of bean collisions; very high
after a shake, and rapidly decaying. If a bean collision occurs, it is simulated by
a burst of exponentially decaying noise. All collision noises pass through a
sharply tuned bandpass lter, which simulates the resonance of the gourd.
Assessment of Physical Models of Particles
Many percussion instruments create their distinctive timbre through the accumulation of microsounds. In the case of shaken instruments, this can be
modeled as a stochastic process. Physical models of such instruments produce
granular sounds, either sparse or dense.
My experiences with this technique were based on the MacOS synthesis
program Syd (Bumgardner 1997), which realized Cook's maracas model. It
allowed control of the resonance frequency, the resonance pole, the probability
of bean collisions, the system decay, and the sound decay. Under certain conditions, these sounds can evoke acoustic instruments recorded in an anechoic
environment. When the range of parameter settings is extended, sounds that are
not particularly realistic, but still interesting are created.
The physical modelling of particles could go much further. As Chapter 3
points out, a vast scientic literature devoted to models of granular processes
has yet to be harnessed for sound synthesis.

Abstract Models of Particles

One can take PhM a step further, and bypass simulations of the acoustic
instrument world. Software allows one to develop models of abstract systems
which can only exist within the memory of the machine. This is a well-established
approach to digital synthesis. Several common synthesis techniques arise out of
ecient computational models only indirectly related to acoustical tone production. The technique of frequency modulation (FM) synthesis, for example,
can readily be seen as an abstract model (Chowning 1973). FM synthesis
is based on a familiar equation in the theory of radio communications. After
much eort, John Chowning tested it in the audio frequency range and then
tuned the synthesis equation to make the results sound somewhat like tradi-

176

Chapter 4

tional instruments. Only when the equation is heavily constrained, however,

does it resemble a recognizable instrument, otherwise it produces abstract
``radiosonic'' timbres.
Abstract models can also control particle synthesis. Chapter 3 described
control schemes for granular synthesis based on recursive and chaotic algorithms. The next section discusses an even more abstract approach, where the
particles are part of the model, and the sound is a side-eect of their interactions.
Abstract Particle Synthesis
To sonify is to translate data into sound. Abstract models sonify particle interactions. Thus the abstract model links particle interactions to sound generation.
The sound produced by an abstract model may not sound particulated, it
may instead be long and continuous, as in the work of Sturm (1999). Sturm
programmed a model of atomic particles which acted according to the laws
of physics. When the particles interacted in their environment, they created a
unique solution which evolved over time. By using de Broglie's equation
stipulating a relationship between particles and wavesthis simulation produced long sine waves. In theory, any signal can be represented by a unique set
of particles with specic initial conditions and a particular time-variant potential. Sturm hypothesized that granular sounds could be modeled by the motions
of a particle within a box with its interior lined with sound samples.
Xenakis's GENDYN system is another example of the abstract approach
(Xenakis 1992; Serra 1992). Instead of using a model imported from atomic
physics, however, Xenakis invented a nonlinear algorithm for simulating the
behavior of a particle bouncing o elastic barriers. The path traced by the
particle is interpreted directly as a soundwave. The waveform is the trace of
the particle's chaotic path as it bounces back and forth between the barriers
while the model's parameters are changing constantly.
We should also mention the Cosmophone project (Calvet, et al. 2000). The
Cosmophone is a device for detection and sonication of cosmic ray bombardments which works by coupling particle detectors to a real-time data acquisition and sound synthesis system. The sonic output of the system goes to
a multichannel loudspeaker conguration. The information received by the
detectors triggers the emission of sounds. The sounds emitted by the system are
programmable. Clearly, various methods of particle synthesis would be well
worth testing with such a system.

177

Varieties of Particle Synthesis

Table 4.4 Sound particles

Name

Envelope type

Waveform

Characteristics

Grains

Gaussian or arbitrary

Arbitrary, including sampled

Glissons

Gaussian or arbitrary

Arbitrary

Pulsars

(1) Rectangular around

pulsaret, then null;
(2) Gaussian
(3) Expodec
(4) Arbitrary
Expodec or arbitrary

Arbitrary

Gaussian or other, subject to mathematical

constraints
Gaussian or arbitrary
Arbitrary

Sinusoidal

Can be scattered irregularly or metrically; each grain is potentially unique;

Gaussian grains are compatible with
analysis systems like the Gabor transform and the phase vocoder
Variable frequency trajectory; synthesizes glissandi or noise textures
Independent control of fundamental and
formant spectrum; can also be applied in
the infrasonic frequencies as a rhythm
generator; synchronous distribution of
pulsars
Used to synthesize tones; oers independent control of fundamental and formant
spectrum
Particle duration varies with frequency;
starts from an analysis of an existing
source sound
Interdependent synthesis parameters
Flexible graphic design; subject to graphical transformations
Envelope controls formant spectrum
Used for pitched tones; exible control
of spectrum
Synthesizes formants with purely harmonic content
Generally sharp and percussive; each
transwave is unique
Repeats a particle so that it becomes a
continuous tone

Trainlets
Wavelets
Grainlets
Micro-arcs
FOF grains
Vosim grains
Windowfunction pulses
Transient
drawing
Particle cloning

Attack, sustain, release

Linear attack, exponential decay
Rectangular around
pulse then nil
Arbitrary
Arbitrary

Impulse

Sine
Arbitrary, including sampled
Sine
Sine 2 pulses
Blackman-Harris
pulse
Hand-drawn
Arbitrary,
including sampled

178

Chapter 4

Assessment of Abstract Models of Particles

In most implementations, abstract models of particles have been used to produce a pressure wave representing a continuous sound. Most of Sturm's sound
examples sound like the additive synthesis of sine waves, where each sine is
warping or sweeping in frequency, usually in synchrony with the other sines.
Xenakis's compositions using the GENDYN system are quite dierent in sonic
quality from the sound of Sturm's examples: the waveforms are jagged in shape
and rich in spectrum, their behavior is obsessive and erratic, always wavering
and warbling, often explosive. Xenakis's S.709 (1992, EMF CD 003) is a classic
example, meandering wildly like some atomic insect.
In abstract models, the mapping between the particle model and the process
of sound generation may be quite arbitrary. The history of digital sound synthesis contains numerous examples of all manner of algorithmic schemes applied to the control of sound synthesis. Thus one could easily use an abstract
model of particles to control the parameters of an arbitrary synthesis technique;
frequency modulation or granular synthesis, for example.

Summary
All forms of music compositionfrom the freely improvised to the formally
organizedare constrained by their sound materials. The urge to expand the
eld of sound comes from a desire to enrich compositional possibilities, and
much can be gained from the harvest of synthetic waveforms produced by
particle synthesis. In chapter 3 and in this chapter, we have looked at a variety
of sound particles. Chapter 6 describes additional particles derived from windowed spectrum analysis. Table 4.4 summarizes the variety of sound particles
studied in this book.
Artists are frequently recalled to the belief that the splendors of nature surpass anything that human beings can create. The natural world, however, did
not inhibit the rst painters, on the contrary, it inspired them. Similarly, for
composers, the omnipresence of natural sound does nothing to quash the need
to create a virtual sound world. With the sound particles, we cultivate a new
strain of culture within the natural order.

Transformation of Microsound

Synthesis and Transformation

Micromontage
Micromontage in a Graphical Sound Editing and Mixing Program
Micromontage by Script
Micromontage by Algorithmic Process
Composition with Micromontage
Assessment of Micromontage
Granulation
Granulation from Sound Files with Asynchronous Playback
Implementations of Granulation from Sound Files
Selective Granulation
Granulation in Real Time
Assessment of Granulation
The Creatovox Synthesizer: A Brief Description
Synthesis Engine
Musical Interface
Assessment of the Creatovox
Pitch Shifting on a Micro Time Scale
Assessment of Pitch Shifting on a Micro Time Scale
Pitch-Time Changing by Granulation
History of Digital Pitch-Time Changing

180

Chapter 5

Granular Pitch-Time Changing with a Harmonizer

Granular Time Stretching and Shrinking in Cloud Generator
Granular Pitch Shifting: Malah, Jones, and Parks' Methods
Granular Pitch Shifting with Constant Formants: Lent's Method
Granular Pitch Shifting with Variable Formants: Gibson's Method
Filtering on a Micro Time Scale
Experiments with Constant-Q Filters on Grains
Dynamics Processing on a Micro Time Scale
Parameters of Spectral Dynamic Processing
Waveset and Wavecycle Distortions
Waveset Formation
Experiences with Waveset and Wavecycle Distortions
Assessment of Waveset and Wavecycle Distortions
Convolution of Microsounds
Status of Convolution
Impulse Response and Cross-Synthesis
Review of Convolution Theory
Mathematical Denition of Convolution
The Law of Convolution
Relationship of Convolution to Filtering
Fast and Instant Convolution
Real-Time Implementation of Convolution
Musical Signicance of Convolution
Cross-Filtering
Spatiotemporal Eects
Importance of Mixing
Noise Reverberation
Modulation as Convolution
Excitation/Resonance Modeling
Rhythm Input

181

Transformation of Microsound

Convolution and Pulsar Synthesis

Assessment of Convolution with Microsounds
Spatialization of Sound Particles
Virtual Spaces
Scattering Particles in Virtual Spaces
Per-Grain Reverberation
Spatialization by Modulation with Particles
Convolutions with Clouds of Sonic Particles
Eect of Cloud Amplitude Envelope
Eect of Particle Envelope
Eect of Particle Duration
Eect of Particle Frequency
Eect of Window Type in Sectioned Convolution
Particle Pluriphony in Physical Spaces
Assessment of Spatialization on a Micro Time Scale
Summary
Physical laws interrelate all the properties of sound signals; a transformation of
one is bound to aect the others. By the law of convolution, any change to the
time or amplitude structure of a sound simultaneously transforms its spectrum. Reciprocally, to transform the spectrum of a sound is to aect its
temporal morphology. Time, frequency, phase, spectrum, and amplitude are all
interrelated.
The transformation of sound by means of electronic circuits has a distinguished history (Bode 1984). In the 1940s, Pierre Schaeer's work creating
sound eects for radio and lm led to the historic breakthrough known as
musique concrete (Schaeer and Moles 1952). In this new approach to musicmaking, Schaeer used recordings as the material to be organized in a composition. In musique concrete, one of the principal strategies of variation, and
therefore of composition itself is the transformation of the source material.
Transformation has now become a central aspect of our sonic art, taking place
on many time scales, from global modications of macrostructure to micro
operations on individual particles and samples. Temporal transformations alter
the time scale of the original sound through stretching, shrinking, or reversal.

182

Chapter 5

Other operations smear, scramble, or warp the temporal morphology through

convolution, granulation, or modulation.
This chapter examines transformations based on microtemporal representations of sound. It begins by looking at frame-based audio data coding, micromontage, and granulation. After granulating a sound, one can make many
transformations on a micro scale, including pitch-shifting, pitch-time changing,
ltering, dynamics processing, and reverberation. After a presentation of
waveset transformations, we introduce the theory of convolution, and lastly the
spatialization of sound particles.

Synthesis and Transformation

Synthesis of sound and transformation of sound have always been closely related. To impose an amplitude envelope on a sine wavea basic technique in
synthesisis to transform the signal: sharp or smooth attack, single or multiple
iterations, steady or tremolo sustain, legato or staccato decay. How then can we
distinguish between synthesis techniques and the transformations presented in
this chapter and in chapter 6?
In general, synthesis begins with raw material on a micro time scale: a single
period of a waveform, an impulse, or a burst of uniformly distributed noise.
Transformations, on the contrary, usually begin with a relatively long stream of
audio samples (greater than 100 ms) on the sound object time scale. Transformations apply to a vast realm of recorded or sampled sounds, including
traditional acoustic instruments, electronic instruments, environmental and industrial noises, the cries of birds and animals, insect noises, and the sounds of
human beings. The samples to be transformed can be extracted from a sound
le or read directly from a real-time source. Often, we can identify the source of
microphone-collected samples, and the goal of the transformation may involve
preserving this source identity.
The transformations described in this chapter do not involve a stage of windowed spectrum analysis, which, instead, are surveyed in chapter 6.

Micromontage
Certain transformations clearly articulate the granular texture of sound. Micromontage extracts particles from sound les and rearranges them. The term

183

Transformation of Microsound

Figure 5.1 Micromontage shown as a display of one hundred and thirty-six sound les
organized in a graphical mixing program. The les appear in twelve tracks stacked vertically, while time moves from left to right. Notice that the total duration of this phrase
is 3.7 seconds.

``montage'' derives from cinema, where it refers to a sequence of rapid images

connected through cutting, splicing, dissolving, and other editing operations.
With micromontage, a composer can position each particle precisely in time,
constructing complex sound patterns by assembling dozens of smaller sounds.
Micromontage can be realized by several means: graphical editing, typing a
script, or writing a computer program to automate the process.
Micromontage in a Graphical Sound Editing and Mixing Program
A graphical sound editing and mixing program provides a two-dimensional
view of sonic structure (gure 5.1). The vertical axis presents multiple rows of
tracks, while the horizontal axis presents a time line. An early sound mixing
program with a horizontal time-line graphical interface was MacMix (Freed

184

Chapter 5

1987). This interface is now standard in audio workstations, including such

popular programs as Digidesign's Pro Tools and many others.
Micromontage requires the following steps:
1 Assemble a library of soundles, including sounds on several dierent time
scales
1 Import the les into the library of the editing and mixing program
1 Use the cursor to position each sound at a specic time-point in a track.
Most programs allow for multiple positionings, so that a single le can be
pasted in repeatedly
1 Edit the duration, amplitude, and spatial position of the sounds on all time
scales
Each microsound positioned within a track is like a sonic brush stroke, and
as in a painting, it may take thousands of strokes to complete the work.
Graphical sound editing and mixing programs oer a multiscale perspective.
One can view the innermost details of sonic material, permitting microsurgery
on individual sample points. Zooming out to the time scale of objects, one can
edit the envelope of a sound until it has just the right weight and shape within
a phrase. Zooming out still further one can shape large sound blocks and
rearrange macrostructure. The availability of dozens of tracks lets composers
work precisely on the micro time scale. A set of sounds that cannot t on a
single track can overlap on several tracks. Multitrack mixing capability also facilitates the construction of dense and variegated textures.
Micromontage by Script
Micromontage by script emerged in the 1980s, when dialects of the Music N
synthesis languages were extended to handle input from sampled sound les. In
the Csound language developed at MIT, the soundin and loscil unit generators
read from a sampled input le (Boulanger 2000). Such a capability made it
possible to create arbitrarily complex montages of sound les according to a
precise numerical score or script. In these languages, a note statement creates a
sound event. In the case of micromontage, each note statement corresponds to
a microevent. Thus one must stipulate hundreds or thousands of microevents to
create a composition. This process is laborious, but it oers extraordinary
control over the montage.

185

Transformation of Microsound

Micromontage by Algorithmic Process

This technique has similarities to micromontage by script. Instead of specifying
each particle manually, however, a computer program handles their extraction
and time-scattering. The program reads a high-level specication of the montage from the composer. It then generates all the note statements needed to
realize the montage. The high-level specications might include a list of sound
les, the location and length of the segments to be extracted, the overall tendencies of the montage, and so on.
Here is a small excerpt of a micromontage text produced by a program that I
wrote in 1993. It has the format of a typical note list.
Start

Duration

Soundle

Amplitude

Location

0.137
0.281
0.346
0.628
0.748
0.847
0.974
...

0.136
0.164
0.132
0.121
0.174
0.062
0.154

8
10
12
1
3
6
8

0.742
0.733
0.729
0.711
0.693
0.687
0.686

0.985
0.899
0.721
0.178
0.555
0.159
0.031

Seven events occur in the interval of one second. Each sound le has a number
labelfrom 1 to 12 in this caseshown in the third column. The location
parameter indicates spatial location in the stereo eld, with 1 corresponding
to left, and 0 corresponding to right. The Granulate program generates the
score for the micromontage according to high-level instructions stipulated by
the composer. The composer species the parameters of a cloud, such as which
sound les to granulate, the density of particles, their amplitude, shape, and so
on. Each cloud may contain hundreds of particles.
The eect of automated micromontage is much the same as granulation. One
dierence between them is that micromontage is based on a scripta text
that is read by a Music N synthesis program, meaning that the user can edit the
script before the montage is rendered into sound.
Composition with Micromontage
Micromontage has been a specialty of the composer Horacio Vaggione for
some time, in such works as Octuor (1982), Thema (1985, Wergo 2026-2), and

186

Chapter 5

Schall (1995, Mnemosyne Musique Media LDC 278-1102). In Octuor, the

composer began by synthesizing a set of ve sound les. He then segmented
them into small fragments using the S sound editor, and mixed the fragments
into medium-scale and large-scale structures using the MIX program (Moorer
1977b). He further processed these structures and combined them in eight-voice
polyphony using an automated mixing program (Vaggione 1984).
Thema explores automated mixing in the Cmusic language (Moore 1990). It
features streams of microelementssuch as resonant bass saxophone breathburstsscattered in both synchronous and asynchronous patterns along the
time line.
Figure 5.2 shows time-domain (a) and frequency-domain (b) views of a 120-ms
frame, which partitions into ve microsections A-E. Microsection A contains 2.5
cycles of a waveform in a space of 34 ms, corresponding to a low fundamental
frequency of 73 Hz. Microsection B begins with an impulse of 90 msec pulsewidth, corresponding to a fundamental of 1.1 kHz, followed by decaying bursts
in a pattern: 17 ms, 8 ms, 17 ms, 8 ms. Superimposed on these bursts is a waveform with a period of 589 msec, corresponding to a fundamental of 1.697 Hz.
Microsection C lasts 21 ms divided into 2.5 periods, corresponding to a low-frequency fundamental of 116 Hz. Microsection D contains a single 113 msec click,
corresponding to a frequency of 8.894 kHz. Finally, microsection E lasts 13 ms
and contains two periods of the frequency 153.8 Hz (slightly south of D-sharp),
upon which are superimposed partials in the 8 kHz region.
The raw material of Schall consists of sampled piano sounds, granulated and
transformed by convolution, waveshaping, and the phase vocoder. According
to the composer:
The work plays essentially with tiny textures of feeble intensity, composed of multiple
strata, which contrast with some stronger objects of dierent sizes, in a kind of dialog
between the near and the faras an expression of a concern with a detailed articulation
of sound objects at dierent time scales. (Vaggione 1995 program notes)

For more about this composition, see chapter 7, which also discusses the
work of the Princeton-based composer Paul Lansky, another pioneer of
micromontage.
Assessment of Micromontage
Micromontage is an open-ended approach, still with many unexploited aesthetic possibilities. Granulation techniques, presented next, have absorbed

187

Transformation of Microsound

Figure 5.2 120-ms excerpt from H. Vaggione's Thema (1985). (a) Time-domain view
shows microtemporal variations divided into ve sections A-E. See the text for an explanation. The amplitude of the signal has been compressed in order to highlight lowlevel microvariations. (b) Frequency-domain view plotted on a linear frequency scale.

many of the techniques of micromontage. Perhaps the best way to draw a distinction between granulation and micromontage is to observe that granulation
is inevitably an automatic process, whereas a sound-artist can realize micromontage by working directly, point by point. It therefore demands unusual
patience.

Granulation
The automatic granulation of sampled sounds is a powerful technique for
sound transformation. To granulate means to segment a sound signal into tiny

188

Chapter 5

grains. (The French term brassage, referring to ``mashing'' or ``mixing,'' is also

used.) The sound may be further modied after this segmentation. The granulation algorithm then reassembles the grains in a new time order and microrhythm. The precise manner in which this occurs varies from program to
program.
Granulation is a purely time-domain operation. This stands in contrast to
techniques such as the Fourier, wavelet, and Gabor transforms, which analyze
a signal in both time and frequency. These techniques also granulate, but they
then apply spectrum analysis to each grain. In spectrum analysis, the initial
stage of granulation is called ``windowing.'' (See chapter 6.)
The parameters of granulation include the following:
1 Selection orderfrom the input stream: sequential (left to right), quasisequential, random (unordered)
1 Pitch transposition of the grains
1 Amplitude of the grains
1 Spatial position of the grains
1 Spatial trajectory of the grains (eective only on large grains)
1 Grain duration
1 Grain densitynumber of grains per second
1 Grain envelope shape
1 Temporal patternsynchronous or asynchronous
1 Signal processing eects applied on a grain-by-grain basislters, reverberators, etc.
Granulation parameters can be controlled by a numerical script, by graphical
envelope editing, by physical controllers, or by deterministic or stochastic
algorithms. Since granulation accepts any input sound, a truly vast range of
sonic textures can emerge from a granulator. The density parameter is key, as it
determines the overall texture, whether empty, sparse, gurgling, or continuous.
Granulation from Sound Files with Asynchronous Playback
Asynchronous playback from a sound le allows much more freedom than realtime granulation of a sound coming from a microphone. Because the sound to
be granulated resides in a le, we have the exibility of extracting individual
grains in any order: sequential, reversed, random, statistical evolution, or in

189

Transformation of Microsound

an arbitrary succession. For example, we can extract a single large grain from
a snare drum and clone a periodic sequence of hundreds of grains to create
a single-stroke roll. To avoid the repetitious quality of commercial drum
machines and samplers, a variation on this method is to select grains from
several strokes of a roll and to extract each grain from a dierent set of samples
in the selected sound le.
One can liken granulation to scattering sound particles with a precision spray
jet of sampled sound waveforms. Grains sampled from dierent instruments
can be mingled, and grains taken from several sound les can create interwoven
fabrics of sound. When the sound les consist of dierent notes of a scale, the
result is a ringing harmonic cloud, where the layers stratify at particular pitches.
When two harmonic clouds overlap, the sound is a statistical evolution from
the rst cloud to the second.
It is fascinating to experiment with extracting grains from several dierent
sources, leading to hybrid textures, such as grains from a cello or voice mixed
with grains from a cymbal. By controlling the distribution of grains from different sources, we can create clouds which evolve from one texture to another.
In any granulation method, grain duration is an important parameter. As
chapter 3 points out, the timbre of a grain correlates strongly with its duration.
When the duration of the grain is very short (<40 ms), sampled sound les lose
their intrinsic identiable qualities. For sound les consisting of spoken text
or other identiable material that is to be preserved, longer duration grains
(>40 ms) work better.
Implementations of Granulation from Sound Files
Chapters 3 and 7 describe implementations of granular synthesis with synthetic
waveforms. Granulation, on the other hand, implies the existence of analog-todigital converters and software for sound recording and editing.
In 1981, I was working at the MIT Experimental Music Studio. At the time,
we referred to analog-to-digital conversion by the now quaint term ``digitizing.''
The software for digitization and editing was primitive, written by students in
short-term research projects. In this fragile environment, I managed to carry
out a number of experiments in sound le granulation using a program that I
wrote in the C language. This program read a script that I prepared, which
described the parameters of a granulated sound cloud. It generated a score
consisting of hundreds of note statements in the Music 11 language. I then ran
Music 11 with a simple instrument based on the soundin unit generator. The

190

Chapter 5

instrument extracted grains from the sample les, applied a quasi-Gaussian

envelope to them at a certain amplitude, and wrote each grain to a particular
point in the four-channel output sound le.
The primary source material consisted of a collection of monaural percussion
sound les digitized by my colleague Christopher Fry, and also some saxophone samples that I digitized (Roads 1985b, 1985c). By extracting grains from
these les, I created eects such as cymbal and drum rolls, as well as rolls
derived from the grains of alto saxophone tones. Since the studio was quadraphonic, I could send the rolls around the studio.
Working at Simon Fraser University, Barry Truax was the rst to implement
real-time granulation of a sampled sound (Truax 1986, 1987). The GSAMX
program operated on a brief recorded sample, no longer than 170 ms. This
limitation was due to the very small memory of the sound processor, a Digital
Music Systems DMX-1000 with 4 kwords of memory. This modest conguration nonetheless produced up to twenty simultaneous granular streams, with
each stream containing up to one hundred grains per second, resulting in an
impressive throughput of two thousand grains per second (Truax 1990a).
In 1990, Truax improved his system to allow extended recording of 16-bit
samples to hard disk, with granulation and time stretching. This conguration
enabled him to realize a number of pioneering compositions (see chapter 7).
By the late 1980s, it was possible to sample sounds on a personal computer
using commercial hardware and software. During a 1991 residency at the
Kunitachi College of Music in Tokyo, I wrote the Granulate program in the C
language. Designed for the synthesis of complex multicolored textures, Granulate used the engine of Csound to pulverize and scatter up to sixty-four sound
les.
Working in Paris at Les Ateliers UPIC in 1995, John Alexander and I
developed the Cloud Generator program, which granulates a single sound le
at a time (Roads and Alexander 1995). Appendix A documents this program.
Most of my current work with granulation involves programs written in the
SuperCollider 2 language, which execute in real time (McCartney 1998).
Many commercial and shareware granulation packages are now available,
even excluding those designed solely for time-shrinking and stretching. A typical example is the SoundMagic FX package by Michael Norris (1997), which
runs on the MacOS platform. Sound FX includes operations labeled ``brassage
time-stretch,'' ``sample hose,'' and ``chunk munger'' for granulation with multiple eects.

191

Transformation of Microsound

Selective Granulation
An increasing number of signal processing operations analyze and separate
dierent components of a sound signal. One set of components may be retained, while the other is discarded. Alternatively, the components can be treated
in dierent ways and then recombined. The listener recognizes the identity of
the original sound, but with an interesting variation in its components.
What are some of the ways of separating a sound into components? One is to
send the sound through a bank of lters, where each lter tunes to a dierent
center frequency. This results in a number of output signals that dier in their
frequency band. Another technique is to set an amplitude threshold and separate the low level and high level parts of a signal. A third way is to separate the
chaotic excitation and the stable resonance parts of a signal. Sounds can be separated by spatial position, by duration, by attack shape, and so on. Generally,
there are no limits on the number of ways in which a given sound can be divided.
Selective granulation means granulating one of these separated components.
For example, only those sounds that fall above or below a given amplitude
threshold. This principle is the foundation of dynamics processing on a micro
time scale (see the description later). The granulated sounds can then be combined with the original to create a new hybrid texture.
Granulation in Real Time
Granulation in real time takes two forms. The rst is the granulation of an incoming sound source, such as the signal coming from a microphone. The second
is the granulation of stored sound les with real-time control. The granulator
can be thought of as a delay line. As the sound passes through the delay line, the
granulator sets and moves various pointers which identify memory locations in
the delay line and extract grains. Several eects are possible:
1. Looping through the incoming samples repeatedly, causes the incoming
sound to be time-stretched. The playback or looping rate can vary by
changing the speed at which the granulator reads through the samples: from
normal speed to a slowed-down rate in which a single grain repeats over and
over. Here a brief grain telescopes into a slowly evolving sound object that
lasts hundreds of times longer than the original.
2. Overlaying many copies of a grain with dierent phase delays increases its
perceived volume and creates a kind of chorus eect.

192

Chapter 5

3. Varying the size and shape of the granulation window introduces amplitude
modulation with its noticeable spectral productsdistorting the input sound
in a controllable way.
Barry Truax was the pioneer of real-time granulation and is its most inveterate exponent (Truax 1986, 1987, 1988, 1990a, 1990b, 1992, 1994a, 1994b,
1995, 1996a, 1996b; Keller and Truax 1998; see also chapters 3 and 7). Numerous programs now oer granulation in real time, as chapter 3 describes.
Granulation of incoming signals is limited by the forward direction of the
arrow of time. As Truax stated:
Granulation may be many things, but it is not omniscient.

(Truax 1994a)

In other words, the read pointer in a real-time system can never peek into the
future. This eliminates the possibility of time-shrinking in real time. In contrast,
the other form of granulation, reading from stored sound les, has the advantage that it can select grains from any time-point in the sound, making it
possible to time shrink and time scramble. Regardless of whether the source
is coming in ``live'' or from a stored sound le, real-time control is especially
eective in creating an unending collection of variants of a given sound.
Assessment of Granulation
Granulating and powdering are, strictly speaking, nothing other than mechanical . . . operations, the object of which is to separate the molecules of a material and to reduce them to
very ne particles. But . . . they cannot reach the level of the internal structure of the material. . . . Thus every molecule, after granulation, still resembles the original material. This
contrasts with the true chemical operations, such as dissolution, which change intimately
the structure of the material. (Antoine Lavoisier 1789, quoted in Vaggione 1996b)

Granulation reduces a continuous sound to a ux of particles. As the great

French chemist Lavoisier observed, molecules of grains resemble molecules of
the original material. Yet the process of granulation alters their arrangement.
This alteration may be subtle, such as a spatial displacement of the grains that
leaves the source sound otherwise intact. On the other hand, just as a brilliant
multicolored stone may be reduced to an indistinct grey powder, certain granulations render the source unrecognizable by destroying its morphology. The
next chapter deals with ``chemical'' changes in sound composition. These entail
alterations in the spectral composition of the sound molecules.
For many years, granulation techniques were available only to those few
composers who had implemented them. As new implementations have been

193

Transformation of Microsound

made available to a larger circle of musicians, granulation has proven to be

popular. To cite an example, at the 1998 Brazilian Computer Music Symposium in the opening concert (featuring composers from Brazil, Italy, Germany,
England, and Taiwan) eight out of ten works played used granulation.
With the spread of granulation technology, musical experience has also
deepened. Naturally more is known about how to compose with granulated
sounds than when they were rst invented and recent works show considerable
renement in their manipulation of granulation parameters. Chapter 7 surveys
this scene.

The Creatovox Synthesizer: A Brief Description

In the early 1990s, I began a design notebook containing schemes for real-time
granular synthesis, including both scheduling algorithms as well as protocols
for musical interaction (Roads 19921997, 1998b). It seemed to me that the
technology had reached a point where it would be possible to build a low-cost
synthesizer for granular synthesis. This instrument might look like a standard
keyboard, with the addition of a few special controllers and could be compatible with the MIDI protocol. As a guiding principle, I wanted to convey ``an
illusion of simplicity'' to the performing musician by hiding unnecessary technical details.
My original plan was to develop a prototype system based around a dedicated digital signal processor (DSP) chip. It became clear that this approach
would require a great deal of time and money. In 1986, a major synthesizer
manufacturer solicited my opinion on promising synthesis techniques and I
suggested that a granular synthesizer would be a good direction. I did not hear
from them again. In 1993, representatives from another synthesizer manufacturer solicited my opinion. I again suggested a granular synthesizer, but apparently, they decided to do something less interesting. The synthesizer project
languished.
By the late 1990s, general-purpose microprocessors such as the Apple G3
became fast enough to handle multiple voices of particle synthesis in real time.
In 1998, James McCartney released the SuperCollider 2 language (McCartney
1998) which combined musical interaction with synthesis and sound processing,
and seemed ideal for prototyping the granular synthesizer. I launched the
Creatovox research project at the Center for Research in Electronic Art Technology (CREATE), University of California, Santa Barbara building upon

194

Chapter 5

there developments. Its goal was to invent a prototype instrument optimized

specically for expressive performance of particle synthesis and granulation.
The design of the Creatovox has two components: the synthesis engine
how to schedule events, and how to synthesize and process the sound, and the
musical performance interfacethe controllers and displays manipulated by
the musicianand their protocol. The account here is cursory, as the Creatovox continues to evolve.
Synthesis Engine
Many technical issues in the design of the Creatovox's synthesis engine have to
do with particle scheduling, setting limits on particle overlap, balancing the
amplitudes of the particles, and determining what to do when the performer
attempts to exceed computation limits. In this last case, the goal is to degrade
performance as gracefully as possible. To generate sucient grain density, the
synthesis engine must be capable of realizing multiple particle streams simultaneously, with individual envelopes and waveforms for each particle.
Current microprocessors have enough speed to implement signal processing
operations that transform the particle streams. Fast convolution, for example,
allows for cross-synthesis, morphing, and granular reverberation eects (Roads
1993a, 1997).
The prototype Creatovox realizes a variety of particle synthesis techniques,
including pure granular synthesis with sinusoidal waveforms and Gaussian
envelopes, granular synthesis with other waveforms and envelopes, grainlet and
glisson synthesis, trainlet synthesis, pulsar synthesis, and granulation of sampled sound les. These can be combined with real-time signal processing eects
running on the same microprocessor.
Musical Interface
The Creatovox is played using standard MIDI keyboard controllers equipped
with pitch bend and modulation wheels. To this is added a pedal board, expression (volume) pedal, sustain and sostenuto footswitches, MIDI joysticks, and a
fader box. Joysticks are ecient controllers as they let the performer control
three separate parameters (left-right, forward-backward, up-down) with one
gesture. The mapping between the various controllers and the synthesizer engine, with all the tuning this requires, constitutes a signicant portion of our
research, and it is here that we have most rened the playability and virtuoso
potential of the instrument.

195

Transformation of Microsound

Assessment of the Creatovox

The Creatovox emitted its birth cries on 13 July 1999 in CREATE's Studio
Theremin, in the presence of a project team consisting of Alberto de Campo,
Ching-Wei Chen, and myself. As of January 2000, another round of development (coded by Alberto de Campo) had produced octophonic output. Demonstrated for the rst time at the CREATE symposium on Sound in Space in
March 2000, it continues to be the subject of ongoing renement and musical
experimentation (gure 5.3). Its sonic palette is wide, yet its signature is quite
distinct, even at rst playing. The rst composer to use the Creatovox was Bebe
Barron, in the summer of 2000. Her fascinating piece, Mixed Emotions, was the
result.

Pitch Shifting on a Micro Time Scale

Pitch shifting means transposing the fundamental frequency of an audio signal.
(Not all sounds, however, have a fundamental frequency.) In the simplest case,
this means changing the playback rate. Faster playback shifts the pitch up and
shortens the sound, while slower playback shifts the pitch down and lengthens
it. (Another eect called pitch-time changing preserves the original duration.
The next section looks at this eect.) In the classic electronic music studios of
the past, the primary device for pitch shifting was the varispeed reel-to-reel tape
recorder. By turning a knob, one could speed up or slow down the playback
motor. It took several seconds for the tape recorder to go from one speed to
another, but this slow change was sometimes a desired part of the eect.
Digital techniques let us alter the playback speed instantaneously, over any
time scale. It is possible, for example, to pitch shift to dierent intervals on a
grain-by-grain basis. The main parameters being the amount of pitch shift (in
percent, semitones, or cents, where a 5.9% pitch shift represents one semitone or
100 cents), grain duration, and grain overlap.
Assessment of Pitch Shifting on a Micro Time Scale
The perceptual eect of pitch shifting on a micro time scale depends on the
duration of the grains and the frequency variance of the pitch shift. Large
grains and large variances result in melodic elaborations of the source. Short

196

Chapter 5

Figure 5.3 The author playing the Creatovox synthesizer in CREATE's Studio Varese,
January 2000. The physical hardware consists of a MIDI keyboard controller, MIDI
joystick, and footpedals. A Power Macintosh computer receives the MIDI data. The
Creatovox software interprets the MIDI data and synthesizes the sound, with octophonic output through a Digidesign Pro Tools 24 converter.

197

Transformation of Microsound

grains and large variances create noise bands. Large grains and small variances,
when combined with overlapping grains, produce a multiple-voice chorus
eect. When the grain durations go below about 50 ms, the chorus turns into
a ghostly whisper, as the pitch variations become microscopic and the signal
dissolves into noise.

Pitch-Time Changing by Granulation

The eect described above, simple pitch transposition, changes the playback
sampling rate, with a corresponding change in the duration of the sound.
Another eect, pitch-time changing, has been applied extensively in computer
music. Pitch-time changing presents two facets. We can stretch the duration of
a sound (time expansion) or shrink it (time compression), while the pitch remains unchanged. Alternatively, we can transpose the pitch of the sound up or
down, while the duration remains constant. These eects can be achieved, with
varying degrees of success, through various strategies. Here we concentrate on
time-domain granulation techniques.
To double the duration of a sampled signal, the algorithm segments it into
grains, cloning each so that two grains appear for every one in the original. To
halve the duration, it deletes every other grain. The grain preserves the local
frequency content, while the algorithm alters the time scale by cloning (to
stretch duration) or deleting (to shrink duration) grains. To shift the pitch of a
sampled signal up an octave but not change its duration, the algorithm doubles
the playback sampling rate while cloning every grain to restore the duration to
the original. To shift the pitch down an octave but not change the duration, the
algorithm halves the playback sampling rate and deletes every other grain to
restore the duration to the original length.
These operations double or halve pitch or time, but such operations are not
limited to products of two. The frequency and time scale can be altered by
arbitrary ratios, by sample rate changing with grain cloning or deleting in
corresponding ratios.
Some of the most eective time-stretching happens when applied in a selective, context-sensitive manner. To preserve the identity of an original sound,
one might process only the steady-state part of the signal, and not the transient
attack. In stretching speech, for example, intelligibility and ``naturalness'' can
be enhanced by stretching vowels more than consonants.

198

Chapter 5

History of Digital Pitch-Time Changing

Chapter 2 described the pioneering role played by Gabor and others in developing sound granulators based on electro-optical technology. This section
describes the rst attempts to granulate and pitch-time change sound using
digital technology.
Pioneering research at the University of Illinois at Urbana-Champaign led to
an early digital implementation of pitch-time changing (Otis, Grossman, and
Cuomo 1968). This research simulated the eect of rotating-head sampling, it
also made plain the aws in this method. The main problem was that the
waveforms at the beginning and end of a sampled grain did not always match
in level with preceding and successive resampled grains, creating a transient at
the junction of the two grains. Electromechanical time-granulators and some
digital implementations exhibited a periodic clicking sound caused by these
splicing transients.
Lee (1972) developed the Lexicon Varispeech system as a digital time compressor/expander interfaced with an analog cassette recorder. This design featured an electronic circuit for level matching at splice points to reduce clicking.
Scott and Gerber (1972) developed an improved digital version of the Springer
Tempophon that used rectangular windows but took pitch into account to determine the optimum window size.
Granular Pitch-Time Changing with a Harmonizer
A harmonizer is a real-time transposing device which shifts the pitch of an
incoming signal without altering its duration. Based purely on time-domain
techniques, the Eventide H910 Harmonizer, released in the mid-1970s, was the
rst commercially available digital device of this type (Bode 1984).
The following passage describes the Publison, a sampling eects processor
developed in France in the early 1980s. The basic notion of a harmonizer is to
load a random-access memory with an incoming signal sampled at a rate of
SRin and to read out the samples at a rate SRout . The ratio SRin =SRout determines the pitch change. To maintain a continuous output signal, the machine
must repeat samples (for upward pitch shifts) or skip samples (for downward
pitch shifts). Because the output address pointer repeatedly overtakes the input
address pointer (for pitch increases) or is overtaken by the recirculating input
address pointer (for pitch decreases), the output address must occasionally
jump to a new point in the memory. To make this splice inaudible, the precise

199

Transformation of Microsound

size of the jump derives from an estimate of the periodicity (pitch) of the incoming signal. When the harmonizer decides to splice, a smoothing fade-out
envelope ramps the amplitude of the presplice signal to zero and a corresponding fade-in envelope ramps to postsplice the signal to full amplitude.
Renements can be added to this basic scheme to improve the audio quality.
One is a noise gate connected to the input to the system to ensure that the pitch
shifting does not try to shift any ambient noise associated with the input signal.
The sound quality of a simple harmonizer depends on the nature of the input
signal and on the ratio of pitch change that it is asked to perform. Small pitch
changes tend to generate less audible side eects. Some commercial devices
produce undesirable side eects (such as buzzing at the frequency of the splicing) when used on material such as vocal sounds.
Granular Time Stretching and Shrinking in Cloud Generator
The Cloud Generator program (Roads and Alexander 1995; see the appendix),
can realize a variety of granular time stretching and shrinking eects, some
unique to it. Setting the Selection parameter to ``Granulate'' causes the program to open a le dialog, requesting a stereo le to process. To time-stretch
the le by a factor of two, one sets the cloud duration to be twice the input le's
duration. To time-shrink a sound, one sets the cloud duration to be shorter than
the selected input sound.
The Selection Order parameter applies only to granulated clouds. It determines in what order grains will be selected from the input sound le. Three
options present themselves:
1 RandomThe program selects input grains from random points in the input
sound le.
1 Statistical evolutionThe program selects input grains in a more-or-less leftto-right order, i.e., at the beginning of the cloud there is a high probability
that grains will be selected from the beginning of the input le; at the end of
the cloud there is a high probability that grains will be selected from the end
of the input le.
1 Deterministic progressionThe program selects input grains in a strictly
left-to-right order.
By experimenting with density and grain duration parameters, one obtains a
variety of granular time distortions. With a density of ten grains per second and

200

Chapter 5

a grain duration of 0.1, for example, the time-stretching eect is amplitude

modulated, since the grain envelope is bell-shaped and there is not enough
overlap for a continuous eect. As the density increases to thirty grains per
second, overlap increases. The sound is now continuous but the multiple grains
overlapping cause a comb ltering eect to become prominent. Another interesting eect is created by stretching the sound with short grains, intentionally
leaving gaps between them. The eect can be enhanced by randomizing the
spatial position of each grain.
Granular Pitch Shifting: Malah, Jones, and Parks' Methods
Malah (1979) developed a computer-based pitch-synchronous technique with
triangular windows called time-domain harmonic scaling (TDHS). TDHS achieves pitch synchronization by continuously altering the lengths of the windows
so that signals sum in phase. In a related method, Jones and Parks (1988)
showed how to smoothly reconstruct the signal by using grain envelopes
that overlap slightly, creating a seamless crossfade between grains. In their
approach, the duration of the grains (which varies according to the fundamental frequency) is independent of the crossfade point, which is always in phase
with its successor grain. A local search and correlation over a window no
greater than a 20 ms nds a crossfade point. This calculation operates in real
time.
Granular Pitch Shifting with Constant Formants: Lent's Method
We have already seen how it is possible to shift pitch while preserving the duration of a sound by means of a combination of resampling and granulation. A
problem with this approach is that pitch-shifting changes not only the fundamental frequency but also shifts the spectrum. For vocal sounds with strong
formants, granular pitch shifting by large intervals gives an articial quality.
This articial quality, which has been called the chipmunk eect or munchkinization, was addressed by Lent (1989).
Lent's method was designed to work with strongly pitched sources, such as
the human voice or monophonic instruments. It begins with a pitch estimation
algorithm, in which the signal is rst ltered to remove all but the fundamental
frequency. Every third zero-crossing boundary thus denes a cycle. Once the
pitch period is known, time-shrinking and time-stretching take place. Individual
cycles repeat to stretch the duration of a sound, cycles are deleted to shrink it.

201

Transformation of Microsound

To reduce artefacts, individual cycles are windowed by a smooth Hanning

function which is twice as long as the cycle. These smooth windows overlap
with one another.
Lent introduced a clever trick to this standard and straightforward process.
The pitch-shifting algorithm appends zero-valued samples to the cycle. Thus,
the fundamental frequency of the signal is lowered, while the spectral relationships in the cycle are maintainedpreserving the formants.
Lent does not, however, describe any formant-preserving method for upward
pitch shifting. It should also be noted that the zero-crossing pitch-estimation
method used in Lent's method is easily confused if the input frequency is not
dominated by a fundamental frequency. As our analysis of wavesets later in this
chapter indicates, if the second harmonic is greater than 50% of the amplitude
of the fundamental, additional zero-crossings appear, which could cause errors
in pitch estimation. According to Lent, his method worked well for vocal and
piano tones, where the fundamental is usually strong.
Bristow-Johnson (1995) adapted Lent's method for commercial pitch-shifting
products marketed by the company Wave Mechanics.
Granular Pitch Shifting with Variable Formants: Gibson's Method
Digital audio eects processing has evolved into an important industry, and
commercialization has led to a number of patents for sound granulation
processes. On behalf of Digitech and DOD Electronics, Gibson, Jubien, and
Roden (1996) obtained a patent describing a granular pitch-shifting circuit with
variable formants. A ``vocal harmony processor'' produced a harmonization of
several notes around a singer's fundamental pitch. Gibson's method (as we will
call it) is a minor variation of Lent's, which instead of preserving the formant
structure of a vocal or instrumental tone, allows variations. The method is a
six-stage process:
1. Sample the input signal and store it in memory
2. Determine its pitch
3. Select harmony notes
4. Determine percent of change in sample rate
5. Resample the stored input signal at the desired sampling rate
6. Replicate a portion of the resampled input signal at a rate equal to the fundamental frequency of the desired output signal

202

Chapter 5

The basic dierence between Gibson's and Lent's method is that in step 5 the
input signal can be resampled to change its timbre (either up or down in frequency), and then window-replicated (step 6) with the window length being the
same as the input signal. This produces a tone with an altered timbre but with
the same pitch as the input.
It is also possible to shift the pitch while preserving the formant structure of
the original. If the desired pitch is above the pitch of the input signal by X
semitones, then the algorithm reduces the length of the Hanning window (step
6) by 2X =12 . Alternatively, if the desired pitch is below the input signal by X
semitones, then the algorithm increases the length of the Hanning window by
2 X =12 .

Filtering on a Micro Time Scale

The rst electrical lters, developed in the early part of the twentieth century,
were supratemporal wave-oriented devices, meant to be installed permanently
on telephone lines. The familiar term equalizer originally referred to such a
permanent lter installed in a public address system in order to compensate for
the acoustic deciencies of a hall. Tone controls on radios of the 1940s rolled
o the higher frequencies above 2 kHz. Much later, audio engineers included
lters operating at xed frequencies in audio mixing consoles. By the mid1960s, analog circuit technology allowed for continuously variable or parametric lters, which mixing engineers could control manually by means of
rotary knobs.
In the early days of computer music (before 1990), rudimentary digital lters
appeared in languages such as Music V, Music-11, Csound, etc., but were tricky
to use, owing to their unpredictable output gain and disappointing tone quality.
The technology of digital lter design has greatly improved. Stable constantgain lters are available in real-time synthesis languages, and we can now
extend ltering down to the micro time scale.
My recent experiments have shown that ltering on a micro scale can result
in lively scintillating sonic textures. Microltration couples naturally with granulation. This technique segments a sampled sound into grains, with each grain
processed by a unique lter. The lters articulate microvariations by injecting
spectral coloration into otherwise static textures. In so doing, they foster heterogeneity inside the sonic fabric.

203

Transformation of Microsound

Experiments with Constant-Q Filters on Grains

In 1997, I wrote a program for granulation, in which every extracted grain
passes through a separate lter. This program, GranQ, was written in the
SuperCollider 1 language, an interactive sound processing environment for
MacOS computers (McCartney 1996). GranQ inserts a bandpass or bandreject
lter of variable or constant Q into the granulation routine. Each lter has its
own center frequency and bandwidth, selected randomly within limits stipulated by the user. The number of lters corresponds to the density of grains per
second, which may be a large number (>100/sec).
The most interesting experiments used constant Q lters. The Q of a bandpass lter can be dened as the ratio of the center frequency to the spread of its
3 dB point (cuto point) bandwidth:
Q

fcenter
fhigh flow

where fcenter is the lter's center frequency, fhigh is the upper cuto frequency,
and flow is the lower cuto frequency. Notice that when the center frequency is
constant, adjusting the Q is the same as adjusting the bandwidth. A constant Q
lter, on the other hand, adjusts the bandwidth according to the center frequency, keeping the ratio the same. For example, suppose that we set the Q to
be a constant value of 2. When the center frequency is 250 Hz, the bandwidth is
125 Hz. When the center frequency is 2500 Hz, the bandwidth is 1250 Hz.
Constant Q lters have the advantage that they sculpt the same musical interval
regardless of their center frequency.
GranQ has eleven parameters that greatly aect the output, and which can
be adjusted in real-time as the sound is being processed:
1. Pitchshifts the input to this pitch
2. Pitch Variationamount of random deviation in the pitch
3. Pitch Quantizationrounds the pitch to a multiple of this value, causing
the pitch variation to jump to only a few pitches
4. Time Raterate of scanning through the input sound le
5. Time Dispersionamount of random deviation in the time rate
6. Time Quantizationrounds the time to a multiple of this value, causing
the granulation to scan only a few time points in the le

204

Chapter 5

7. Grain Duration
8. Grain Overlap or Density
9. Amplitude
10. Filter Qconstant over a range of 0.1 to 20
11. Filter Rangeupper and lower limits, from 20 Hz to 20 kHz
When the lter Q and density are high, the granular stream has a liquid
quality. One can disintegrate any sound into a torrent of ltered particles.
GranQ is one of my staple treatments for sound. I applied the program in
creating the second movement of my composition Half-life (1999), the source
material for which consisted of a series of sound les generated by pulsar synthesis (see chapter 4). I processed these les with GranQ at variable densities
and grain sizes, creating the cascading clouds heard in the second movement,
Granules. (See description, chapter 7.)

Dynamics Processing on a Micro Time Scale

Dynamics processing aects the amplitude of audio signals. Such operations
include compression of the envelope amplitude of a sound, limiting, expansion,
and noise gating (Roads 1996). In typical applications, sound engineers apply
these eects on the macro time scale to an entire track or piece. Compression
is common in popular music, lm and video soundtracks, and broadcasts. It
constrains the amplitude of signals to remain within a specic range, allowing
the sound engineer to boost the apparent loudness of the treated signal. The
compression characteristics of a radio or television station may be set once
and then applied to the broadcast signal for years at a time. This is an extreme example of the supratemporal wave-oriented approach to musical signal
processing.
Digital techniques let us alter the dynamics of a signal on a short-term basis
in order to articulate microuctuations in amplitude. An example of this type of
processing appears in the SoundHack program (MacOS) under the rubric of
``Spectral Dynamics'' (Erbe 1995). This group of operations applies dynamics
processing to the frames of windowed spectrum analysis (see chapter 6).
Each frame represents a microsound, a time-slice of the Gabor matrix. Spectral
dynamics processing applies dynamics processing (gating, ducking, expansion,
compression) individually to each time-slice and spectral band. Depending on

205

Transformation of Microsound

how the parameters are adjusted, this articulates the microstructure of a given
sound.
Parameters of Spectral Dynamic Processing
The Spectral Dynamics eect in SoundHack oers threshold detection for each
band. This means that one frequency band can have the dynamics process
active, while other bands are inactive. One can select whether to aect sounds
above or below a specied amplitude threshold. One can set the threshold level
to one value for all bands, or to a dierent value for each band by reading in
and analyzing a sound le. The spectrum of this sound le can set the thresholds for each band.
Other parameters let users set the amount of gain or reduction for the bands
that are past a specied threshold. For compression and expansion, it allows
one to set the gain ratio. When aecting sounds below the threshold, the compressor and expander hold the highest level steady and aect lower levels (this is
also known as downward expansion or compression). When aecting sounds
above the threshold, the compressor and expander hold the lower threshold
level steady and compress or expand upwards.
The attack and decay parameters are important for the manipulation of
microsound structure. These set the time window for each band to open or
close. When this value is a large time constant, the algorithm ignores transients.
If the time constant is a small duration, then the eect tends to modulate the
sound le on the time scale of the window within the aected bands.

Waveset and Wavecycle Distortions

Several years ago the British composer Trevor Wishart developed a family of
eccentric sound transformations based on a sound particle that he called a
waveset. He dened a waveset simply as the ``distance from one zero-crossing to
a third zero-crossing'' (Wishart 1994, appendix 2, p. 50). For a pure sinusoidal
signal, a waveset is equivalent to one wave period or wavecycle. For a more
complicated signal with multiple zero-crossings, segmentation by wavesets
breaks the signal into fragments, whose size depends on the signal's partial
structure. One can use the waveset as a unit of sound transformation, sometimes subdividing it into half-wavesets, which span the distance between two
zero crossings.

206

Chapter 5

Figure 5.4 Formation of wavesets. The circles indicate zero crossings. (a) Sine wave
plus second harmonic in a 1 : 0.5 mix. (b) Sine wave plus second harmonic in a 1 : 0.7
mix. (c) Sine wave plus second harmonic in a 1 : 1 mix. (d) Sine wave plus second harmonic in a 0.5 : 1 mix.

Waveset Formation
What types of signals have multiple zero-crossings? Beginning with a sinusoid,
one can add an arbitrary number of partials without crossing zero more than
once within the wave period. The classic sawtooth and square waves are examples of single-cycle waveforms with an arbitrary number of partials, limited
only by the sampling frequency of the synthesis system.
In any sound with a strong fundamental frequency, waveset manipulations are
equivalent to operations on individual cycles. Thus waveset time-stretching produces no artefacts when the signal has a strong and steady fundamental period.
But as gure 5.4 shows, wavesets form when the ratio of the amplitude of the
fundamental to any of the upper partials dips below 1 : 0.5. In speech, multiple
wavesets appear in sounds such as whispering, where the fundamental drops out.
Experiences with Waveset and Wavecycle Distortions
Table 5.1 summarizes Wishart's catalog of distortions based on wavesets and
wavecycles. The primary documentation of these distortions is a series of

207

Transformation of Microsound
Table 5.1 Waveset and wavecycle transformations in the Composer's Desktop Project
software
Waveset transformations
Waveset transposition
Waveset reversal
Waveset shaking
Waveset inversion
Waveset omission
Waveset shuing
Waveset distortion
Waveset substitution
Waveset harmonic
distortion
Waveset averaging

Waveset enveloping
Waveset transfer
Waveset interleaving,
method 1
Waveset interleaving,
method 2
Waveset time-stretching
Waveset time-shrinking
Waveset normalizing

Substitutes N copies of a waveset in the place of M wavesets,

for example 2 in the space of 1, or 1 in the space of 4, for
doubling and quartering of frequency, respectively
Reverses individual wavesets while retaining their order;
reversals can be of each individual waveset or collections of
N wavesets at a time
Alternates between compressing (in time) and expanding (in
time) successive wavesets
Inverts the phase of all wavesets in a signal; in Wishart's
diagram half-wavesets are inverted
Deletes every Nth waveset, leaving silence in its place; controlled by initial and nal density from 0 to 100%
Permutes collections of wavesets. A simple shue of successive wavesets starting with (a, b, c, d ) becomes (b, a, d, c);
shuing by pairs of wavesets makes (d, c, a, b)
Distorts a signal by squaring and cubing the signal; the example given in Wishart does not indicate how it is tied to waveset boundaries
Replaces wavesets by a stipulated waveform of the same
amplitude, frequency, and time span as the original waveset
Superimposes N harmonics on the waveset fundamental with
a scaling factor M relative to the previous harmonic
Creates a signal containing N new wavesets that are the
average duration and the average amplitude and time
function of the N original wavesets; the overall duration of
the signal is unchanged
Applies an envelope to 1 to N wavesets at a time
Substitutes the waveform of the wavesets in signal A into the
time frames of the wavesets in signal B
Substitutes wavesets from signal A into alternate wavesets of
signal B; applied either to individual wavesets or groups
Interleaves wavesets from signal A with wavesets of signal
B, thus lengthening the output signal; applied either to
individual wavesets or groups
Repeats each waveset N times
Retains only the rst of every N wavesets or retains only the
loudest of every N wavesets
Normalizes every N wavesets above a stipulated amplitude
threshold, thus a 10% threshold has a greater eect than a
90% threshold.

208

Chapter 5
Table 5.1 (continued)
Wavecycle transformations
Distort average
Average the waveshape over N wavecycles
Distort cyclecnt
Count wavecycles in soundle
Distort delete
Time-compress le by deleting wavecycles
Distort divide
Distortion by dividing wavecycle frequency
Distort envel
Impose envelope over each group of N wavecycles
Distort lter
Time-compress sound by ltering out wavecycles
Distort fractal
Superimpose miniature copies of source wavecycles onto
themselves
Distort harmonic
Harmonic distortion by superimposing ``harmonics'' onto
wavecycles
Distort interact
Time-domain interaction of sounds
Distort interpolate
Time-stretch le by repeating wavecycles and interpolating
between them
Distort multiply
Distortion by multiplying wavecycle frequency
Distort omit
Omit A out of every B wavecycles, replacing them by silence
Distort pitch
Pitch warp wavecycles of sound
Distort repeat
Time-stretch le by repeating wavecycles
Distort replace
Strongest wavecycle in each group replaces others
Distort reform
Modify shape of wavecycles
Distort reverse
Cycle-reversal distortion, wavecycles reversed in groups
Distort shue
Distortion by shuing wavecycles
Distort telescope
Time-compress by telescoping N wavecycles to 1

diagrams in his book. The descriptions in table 5.1 are my interpretations of

these sketches.
My own experiments with waveset and wavecycle distortions involved the
SoundMaker editor with its SoundMagic FX plug-in eects (Ricci 1997; Norris
1997). The operations in this package include harmonic distortion, interleaving,
omission, reversal, time stretch, and normalizing. Additional operations apply
to the shorter wavecycles. For example, wavecycle substitution substitutes a
user-stipulated waveform for every wavecycle in the input signal.
Assessment of Waveset and Wavecycle Distortions
Waveset operations are bounded on zero-crossings. This does not exclude disjoint transients in the results; it merely guarantees that the transient does not
occur between waveset boundaries. Certain operations cause transients at every

209

Transformation of Microsound

waveset boundary, so the signal accumulates considerable broadband noise.

The results often resemble the jagged waveforms produced by the waveform
segment synthesis techniques proposed by G. M. Koenig, H. Brun, and P. Berg
(Roads 1996; Berg 1978).
As a unit of sound transformation, a waveset does not seem signicantly
dierent from a wavecycle unless a fundamental frequency is weak or absent.
Wishart justied its use on the following grounds:
The technique is very fast to compute but often introduces strange, signal-dependent
artefacts. (It can therefore be used as a process of constructive distortion in its own
right!) (Wishart 1994, p. 30)

A compact disc supplied with his book contains sound examples of waveset
distortions. The examples of waveset inversion applied to a speaking voice
sound similar to band-shifted radio distortion. Waveset omission creates noiseinfused textures, with temporal gaps as the proportion of omitted wavesets
increases. In waveset substitution, the timbre changes according to the substituted waveform. For example, when sine waves substitute for wavesets, the
result isnot surprisinglymore sinusoidal in quality. Waveset time-stretching, in which each waveset repeats N times, has a distinctly articial quality.
The more extreme waveset distortions destroy the identity of the source signal, turning it into a chain of arbitrary waveform fragments. As the composer
observes, the results are often unpredictable in their detail:
In general, the eects [of waveset time-stretching] produced will not be entirely predictable, but they will be tied to the morphology (time-varying characteristics) of the original
sound. (pp. 401)
Although [waveset averaging] appears to be similar to the process of spectral blurring, it is
in fact quite irrational, averaging the waveset length and the wave shape in perceptually
unpredictable ways. (p. 42)

In general, waveset and wavecycle distortions tend to make a signal sound

more ``electronic,'' and should probably be applied with this in mind.

Convolution of Microsounds
Increased processor speeds make it possible to realize previously exotic and
computationally intensive techniques on personal computers. Convolution is
one such technique. A fundamental operation in signal processing, convolution

210

Chapter 5

``marries'' two signals (Rabiner and Gold 1975). Convolution is also implicit in
signal processing operations such as ltering, modulation, excitation/resonance
modeling, cross-ltering, spatialization, and reverberation. By implementing
these operations as convolutions, we can take them in new and interesting
directions. This section reviews the theory and presents the results of systematic
experimentation with this technique. Throughout it oers practical guidelines
for eective musical use of convolution, and later presents the results of new
applications such as transcription of performed rhythms and convolutions with
sonic particles. Parts of this text derive from Roads (1992b, 1993a, 1996, and
1997).
Status of Convolution
The theory of convolution may remain unfamiliar to most musicians, but to
signal processing engineers it is fundamental: the foundation stone of linear
system theory. Signal processing textbooks often present it tersely, reducing it to
a handful of generalized mathematical cliches. Since these texts are not aimed at
a musically-inclined reader, the audio signicance of convolution is barely
touched upon. Hence engineers are not always aware of the range of convolution
eects in the audio domain (an exception is Dolson and Boulanger 1985).
Listeners are familiar with the eects of convolution, even if they are
unaware of its theory. Convolution may disguise itself under more familiar
terms such as ltering, modulation, and reverberation. Newer software tools
running on personal computers unbundle convolution, oering it as an explicit
operation, and allowing any two sampled les to be convolved (MathWorks
1995; Erbe 1995; Pranger 1999). Such tools provide a stable basis for musical
exploration of convolution, and prompt a need for more universal understanding of its powers. We begin this task here. Those already familiar with the
theory of convolution may want to skip to the section ``Musical Signicance of
Convolution.''
Impulse Response and Cross-Synthesis
The denition of a lter is very broad (Rabiner et al. 1972). Virtually any system that accepts an input signal and emits an output is a lter, and this certainly applies to convolution. A good way to examine the eect of a lter is to
see how it reacts to test signals. One of the most important test signals in signal
processing is the unit impulsean instantaneous burst of energy at maximum
amplitude. In a digital system, the briefest possible signal lasts one sample pe-

211

Transformation of Microsound

riod. Since short-duration signals have broad bandwidths, this signal contains
energy at all frequencies that can be represented at the given sampling frequency. The output signal generated by a lter that is fed a unit impulse is the
impulse response (IR) of the lter. The IR corresponds to the system's amplitude-versus-frequency response (often abbreviated to frequency response). The
IR and the frequency response contain the same informationthe lter's
response to the unit impulsebut plotted in dierent domains. That is, the
IR is a time-domain representation, and the frequency response is a frequencydomain representation.
Convolution serves as the bridge between the time-domain and the frequency-domain. Any lter convolves its impulse response with the input signal to produce a ltered output signal. The implications of convolution in
audio engineering are vast. One can start from the measured IR of any audiofrequency system (microphone, loudspeaker, room, distortion, delay eect,
equalizer, modulator, etc.), and through convolution, impose the characteristics
of this system on any audio signal.
This much is understood in the engineering community. By generalizing the
notion of impulse response, however, one arrives at quite another set of possibilities. Let us consider any sequence of samples as the impulse response of a
hypothetical system. Now we arrive at a new and musically potent application
of convolution: cross-synthesis by convolution of two arbitrary sound signals.
In musical signal processing, the term cross-synthesis describes a number of
dierent techniques that in some way combine the properties of two sounds into
a single sound. This may involve shaping the spectrum, time, or spatial pattern
of one sound by the other.
What then precisely is convolution? The next section presents an intuitive
review of the theory.
Review of Convolution Theory
To understand convolution, let us examine the simplest case: convolution of a
signal a with a unit impulse, which we call unitn. A unit impulse is a digital
sequence dened over n time points. At time n 0, unitn 1, but for all other
values of n, unitn 0. The convolution of an with unitn can be denoted as
follows:
outputn an  unitn an
Here the sign ``'' signies convolution. This results in a set of values for
output that are the same as the original signal an. Thus, convolution with the

212

Chapter 5

unit impulse is said to be an identity operation with respect to convolution, because any function convolved with unitn leaves that function unchanged.
Two other simple cases of convolution tell us enough to predict what will
happen at the sample level with any convolution. If we scale the amplitude of
unitn by a constant c, we can write the operation as follows:
outputn an  c  unitn
The result is simply:
outputn c  an
In other words, we obtain the identity of a, scaled by the constant c.
In the third case, we convolve signal a by a unit impulse that has been timeshifted by t samples. Now the impulse appears at sample n t instead of at
n 0. This can be expressed as follows:
outputn an  unitn t
The result of which is:
outputn an t
That is, output is identical to a except that it is time-shifted by the dierence
between n and t.
Putting together these three cases, we can view any sampled function as a
sequence of scaled and delayed unit impulse functions. They explain the eect
of convolution with any IR. For example, the convolution of any signal a with
another signal b that contains two impulses spaced widely apart results in a
repetition or echo of a starting at the second impulse in b. When the impulses in
b move closer together, the scaled repetitions of b start to overlap.
Thus, to convolve an input sequence an with an arbitrary function bn, we
place a copy of bn at each point of an, scaled by the value of an at that
point. The convolution of a and b is the sum of these scaled and delayed functions. Clearly convolution is not the same as simple multiplication of two signals. The multiplication of one signal a by another signal b means that each
sample of a is multiplied by the corresponding sample in b. Thus:
output1 a1  b1
output2 a2  b2
etc.

213

Transformation of Microsound

In contrast, with convolution each sample of a is multiplied by every sample

of b, creating an array of samples of length b for every sample of a. The convolution is the sum of these arrays. Compare convolution by the unit impulse
(discussed previously) with multiplication by the unit impulse. The multiplication of an by the unit impulse unitn results in all values of outputn being set
to zero except for output0, where unitn equals 1.
Mathematical Denition of Convolution
A mathematical denition of the convolution of two nite sequences of samples
a and b is as follows:
an  bn outputk

N 1
X

an  bn m

m0

where N is the length of the sequence a in samples and m ranges over the entire
length of b. In eect, each sample of an serves as a weighting function for a
delayed copy of bn; these weighted and delayed copies all add together. The
conventional way to calculate this equation is to evaluate the sum for each
value of k. This is direct convolution. At the midpoint of the convolution, n
copies are summed, so the result of this method of convolution is usually
rescaled (i.e., normalized ) afterward.
Convolution lengthens inputs. The length of the output sequence generated
by direct convolution is:
lengthoutput lengtha lengthb 1
In the typical case of an audio lter (lowpass, highpass, bandpass, bandreject),
a is an IR that is very short compared to the length of the b signal. For a
broad smooth lowpass or highpass lter, for example, the IR lasts less than a
millisecond.
The Law of Convolution
A fundamental law of signal processing is that the convolution of two waveforms is equivalent to the multiplication of their spectra. The inverse also holds,
that is, the multiplication of two waveforms is equal to the convolution of their
spectra. Another way of stating this is as follows:
Convolution in the time domain is equal to multiplication in the frequency domain, and vice
versa.

214

Chapter 5

The law of convolution has profound implications. In particular, the convolution of two audio signals is equivalent to ltering the spectrum of one sound
by the spectrum of another sound. Conversely, multiplying two audio signals
(i.e., performing amplitude modulation or ring modulation), is equal to convolving their spectra. Convolution of spectra means that each point in the
discrete frequency spectrum of input a is convolved with every point in the
spectrum of b. Convolution does not distinguish whether its input sequences
represent samples or spectra. To the convolution algorithm they are both just
discrete sequences.
Another implication of the law of convolution is that every time we reshape
the envelope of a sound, we also convolve the spectrum of the envelope with the
spectrum of the reshaped sound. In other words, every time-domain transformation results in a corresponding frequency-domain transformation, and vice versa.
Relationship of Convolution to Filtering
Convolution is directly related to ltering. The equation of a general niteimpulse-response (FIR) lter is as follows:
yn a  xn G b  xn 1 G


i  xn j
We can think of the coecients a; b; . . . i as elements in an array hi, where
each element in hi is multiplied by the corresponding element in array x j.
With this in mind, the general equation of an FIR lter presented earlier can be
restated as a convolution:
yn

N 1
X

hm  xn m

m0

where N is the length of the sequence h in samples, and n ranges over the entire
length of x. Notice that the coecients h play the role of the impulse response
in the convolution equation. And, indeed, the impulse response of an FIR lter
can be taken directly from the value of its coecients. Thus, any FIR lter can
be expressed as a convolution, and vice versa.
Since an innite-impulse response (IIR) lter also convolves, it is reasonable
to ask whether there is also a direct relation between its coecients and its impulse response. In a word, the answer is no. There exist, however, mathematical
techniques that design an IIR lter to approximate a given impulse response.
See Rabiner and Gold (1975).

215

Transformation of Microsound

Fast and Instant Convolution

Direct convolution is notoriously intensive computationally, requiring on the
order of N 2 operations, where N is the length of the longest input sequence.
Thus it is rarely used to implement narrow band lters or reverberators (both of
which have long impulse responses) when simpler methods suce.
Many practical applications of convolution use a method called fast convolution (Stockham 1969). Fast convolution for long sequences takes advantage
of the fact that the product of two N-point Fourier transforms is equal to the
Fourier transform of the convolution of two N-point sequences. Fast convolution means that direct convolution can be replaced by FFTs, which are dramatically quicker for large values of N. In particular, fast convolution takes on
the order of N  log2 N operations. Speedups begin to occur as soon as N
exceeds sixty-four samples. Consider, for example, the direct convolution of a
pair of two-second sounds sampled at 48 kHz. This requires about 96,000 2 or
9.2 billion operations. Fast convolution with the same two sounds requires less
than 1.5 million operations, a speedup by a factor of 6100. Put another way, a
microprocessor that can perform a fast convolution of a pair of two-second
sounds in real time would take three hours and twenty-two minutes to calculate
their direct convolution.
Aldo Piccialli applied the concept of instant convolution to maximize the
eciency of a pitch-synchronous granular analysis/resynthesis system (De Poli
and Piccialli 1991). Instant convolution is a sleight-of-hand that takes advantage of the fact that the convolution of a signal S with a unit impulse is an
identity operation. That is, the convolution of any nite-duration signal S with
a series of impulses reduces to a series of additions of S. In Piccialli's system S
was a grain template that imposed a timbre onto the pitch determined by the
impulse train.
Real-Time Implementation of Convolution
In the block-transform implementation of fast convolution, the spectrum multiplication occurs only after a block of input samples has been collected in
memory and analyzed. Depending on the length of the input signals, this may
involve a signicant delay. Direct convolution has no such delay, but can require an enormous amount of computation. Fortunately for real-time applications where immediate output is needed, it is also possible to implement fast
convolution in sections, that is, a few samples at a time. Sectioned and non-

216

Chapter 5

sectioned convolution generate equivalent results. Rabiner and Gold (1975) and
Kunt (1981) present techniques for sectioned convolution and real-time implementations. Gardner (1995) describes a novel technique that combines direct
and sectioned convolution to eliminate processing delays.
Musical Signicance of Convolution
A veritable catalog of sonic transformations emerges out of convolution: crosslters, spatialization, modulation, models of excitation and resonance, and
time-domain eects. Indeed, some of the most dramatic eects induced by
convolution involve temporal transformations: attack smoothing, multiple
echoes, room simulation, time smearing, and reverberation. The type of eect
achieved depends entirely on the nature of the input signals. Pure convolution
has no control parameters.
The following sections spotlight each type of transformation. A mark (1) in
front of an indented section indicates a practical guideline.
Cross-Filtering
One can implement any lter by convolving an input signal with the impulse
response of the desired lter. In the usual type of FIR audio lter, the IR is
typically less than a few dozen samples in length. The impulse response of a
bandpass lter is precisely a grain with a sinusoidal waveform. The longer the
grain, the stronger is the eect of the lter.
By generalizing the notion of impulse response to include signals of any
length, we enter into the domain of cross-ltering; mapping the time-varying
spectrum envelope of one sound onto another.
1 If both signals are long in duration and one of the input signals has a smooth
attack, the main eect of convolution is a spectrum alteration.
Let us call two sources a and b, and their corresponding analyzed spectra
spectrum_a and spectrum_b. If we multiply each point in spectrum_a with each
corresponding point in spectrum_b and then resynthesize the resulting spectrum, we obtain a time-domain waveform that is the convolution of a with b.
1 If both sources are long duration and each has a strong pitch and one or both
of the sources has a smooth attack, the result will contain both pitches and
the intersection of their spectra.

217

Transformation of Microsound

For example, the convolution of two saxophone tones, each with a smooth attack, mixes their pitches, sounding as though both tones are being played
simultaneously. Unlike simple mixing, however, the ltering eect in convolution accentuates metallic resonances that are common in both tones.
Convolution is particularly sensitive to the attack of its inputs.
1 If either source has a smooth attack, the output will have a smooth attack.
Listening to the results of cross-ltering, one sometimes wishes to increase the
presence of one signal at the expense of the other. Unfortunately, there is no
straightforward way to adjust the ``balance'' of the two sources or to lessen the
convolution eect.
Spatiotemporal Eects
Spatiotemporal eects constitute an important class of transformations induced by convolution. These include such staples as echo, time-smearing, and
reverberation.
Any unit impulse in one of the inputs to the convolution results in a copy of
the other signal. So if we convolve any brief sound with an IR consisting of two
unit impulses spaced one second apart, the result is a clear echo of the rst
sound.
1 To create a multiple echo eect, convolve any sound with a series of impulses
spaced at the desired delay times. For a decaying echo, lower the amplitude
of each successive impulse.
Time-smearing occurs when the pulses in the IR are spaced close together,
causing the convolved copies of the input sound to overlap. If, for example, the
IR consists of a series of twenty impulses spaced 10 ms apart, and the input
sound is 500 ms in duration, then multiple copies of the input sound overlap,
blurring the attack and every other temporal landmark.
The IR of a room contains many impulses, corresponding to reections o
various surfaces of the roomits echo pattern. When such an IR is convolved
with an arbitrary sound, the result is as if that sound had been played in that
room, because it has been mapped into the room's echo pattern.
1 If we convolve sound a with the IR of an acoustic space, and then mix this
convolution with a, the result sounds as if a is within the acoustic space.

218

Chapter 5

We hear reverberation in large churches, concert halls, and other spaces with
high ceilings and reective surfaces. Sounds emitted in these spaces are reinforced by thousands of closely-spaced echoes bouncing o the ceiling, walls,
and oors. Many of these echoes arrive at our ears after reecting o several
surfaces, so we hear them after the original sound has reached our ears; the
myriad echoes fuse into a lingering acoustical ``halo.''
From the point of view of convolution, a reverberator is nothing more than a
particular type of lter with a long IR. Thus we can sample the IR of a reverberant space and then convolve that IR with an input signal. When the convolved sound is mixed with the original sound, the result sounds like the input
signal has been played in the reverberant space.
Importance of Mixing
For realistic spatial eects, it is essential to blend the output of the convolution
with the original signal. In the parlance of reverberation, the convolved output
is the wet (i.e., processed) signal, and the original signal is the dry or unprocessed signal.
1 It is typical to mix the wet signal down 15 dB or more with respect to the
level of the dry signal.
Noise Reverberation
When the peaks in the IR are longer than one sample, the repetitions are timesmeared. The combination of time-smearing and echo explains why an exponentially-decaying noise signal, which contains thousands of sharp peaks in its
attack, results in reverberation eects when convolved with acoustically dry
signals.
1 If the amplitude envelope of a noise signal has a sharp attack and a fast exponential decay, the result of convolution resembles a natural reverberation
envelope.
1 To color this reverberation, one can lter the noise before or after convolving it.
1 If the noise has a slow logarithmic decay gure, the second sound appears to
be suspended in time before the decay.
1 If the noise signal has an exponentially increasing envelope, the second sound
gives the impression of being played in reverse.

219

Transformation of Microsound

Modulation as Convolution
Amplitude and ring modulation (AM and RM) both call for multiplication of
time-domain waveforms. The law of convolution states that multiplication of
two waveforms convolves their spectra. Hence, convolution accounts for the
sidebands that result. Imagine that instead of impulses in the time-domain,
convolution is working on line spectra in the frequency-domain. The same rules
applywith the important dierence that the arithmetic is that of complex
numbers. The FFT, for example, generates a complex number for each spectrum component. Here, the main point is that this representation is symmetric
about 0 Hz, with a replica of each component (halved in amplitude) in the
negative frequency domain. This negative spectrum is rarely plotted, since it has
signicance only inside the FFT. But it helps explain the positive and negative
sidebands generated by AM and RM.
Excitation/Resonance Modeling
Many vocal and instrumental sounds can be simulated by a two-part model: an
excitation signal that is ltered by a resonance. The excitation is a nonlinear
switching action, like the pluck of a string, the buzz of a reed, or a jet of air into
a tube. The resonance is the ltering response of the body of an instrument.
Convolution lets us explore a virtual world in which one sound excites the resonances of another.
Through a careful choice of input signals, convolution can simulate improbable or impossible performance situationsas if one instrument is somehow
playing another. In some cases (e.g., a chain of bells striking a gong), the interaction could be realized in the physical world, others (e.g., a harpsichord
playing a gong), can only be realized in the virtual reality of convolution.
1 To achieve a plausible simulation, the excitation must be a brief, impulse-like
signal, (typically percussive), with a sharp attack (or multiple sharp attacks).
The resonance can be any sound.
Rhythm Input
We have seen that a series of impulses convolved with a brief sound maps that
sound into the time pattern of the impulses. A new application of convolution is
the precise input of performed rhythms. To enter a performed rhythm, one need

220

Chapter 5

only tap with drumsticks on a hard surface, and then convolve those taps with
other sounds.
1 The convolution of a tapped rhythmic pattern with any sound having a sharp
attack causes each tap to be replaced by a copy of the input sound.
This is a direct method of mapping performed rhythms to arbitrary sounds.
Since convolution aligns the sounds to the rhythm with a time resolution of
the sampling rate, this approach is much more precise than a MIDI percussion controller with its temporal resolution of several milliseconds. One can
also layer convolutions using dierent patterns and input sounds. After prepositioning each tap in stereo space, convolution automatically distributes them
spatially.
Convolution and Pulsar Synthesis
The slower the ow of time colors, the greater the clarity with which they can represent
themselves . . . as rhythms. (Koenig 1962)

Chapter 4 discussed pulsar synthesis which generates a stream of microsonic

events at a continuously variable rate, from the audio range down to the
infrasonic frequency domain. Thus it traverses what G. M. Koenig called the
``time-color'' continuum spanning the durations corresponding to timbre
(< @200 msec), pitch (@200 msec to @33 ms), and rhythm (> @33 ms).
The combination of pulsars with convolution is a potent one. Each pulsar,
when convolved with an arbitrary sampled sound, acts as the impulse response
of a lter. The pulsar imposes its temporal prole, spatial position, and spectrum onto the sampled sound. If the same sampled sound convolves with every
pulsar, timbral variations derive from two factors: (1) the ltering eect
imposed by the spectrum of each pulsar, and (2) the time-smearing eects
caused by convolution with pulsar trains whose period is shorter than the
duration of the sampled sound.
This technique requires two collections of sound les: a collection of pulsar
trains, and a collection of sampled sounds. Since convolution is highly sensitive
to attack shape, only objects with sharp attacks preserve the time structure of
precise rhythmic patterns. Long sound objects or those with smooth attacks
lead to time-smearing eects which blur the rhythmic pattern, creating rippling
continua rather than brief time-discrete events.
In pulsar synthesis, the parameters of rhythm, pitch, and timbre can all be
manipulated independently. The technique can generate not only individual

221

Transformation of Microsound

sound objects but also rhythmic sequencesallowing sound composition to

pass directly from microstructure to mesostructure.
Assessment of Convolution with Microsounds
It is likely that convolution techniques will play an increasing role in sound
transformation, particularly as real-time convolvers such as spatializers and
reverberators become more generally available.
A word of caution is in order: Convolution is a powerful transformer, and
when applied to two arbitrary input signals, it can easily destroy the identity of
both sources. Many convolutions that rst appear to be interesting musical
ideas can result in amorphous sound blobs. An understanding of a convolution's musical implications will save the composer much time spent searching
for interesting results.

Spatialization of Sound Particles

When a small box s is situated, relatively at rest, inside a hollow space of a larger box S,
then the hollow space of s is a part of the hollow space of S, and the same ``space,'' which
contains both of them, belongs to each of these boxes. When s is in motion relative to S,
however, the concept is less simple. One is inclined to think that s encloses always the same
space, but a variable part of S. It then becomes necessary to apportion to each box its
particular space, not thought of as bounded, and to assume that these two spaces are in
motion with respect to each other. Before one has become aware of this complication, space
appears as an unbounded medium or container in which material objects swim about. But it
must now be remembered that there is an innite number of spaces in motion with respect
to one another. The concept of space as something existing objectively and independent of
things belongs to prescientic thought. . . . The subtlety of the concept of space was
enhanced by the discovery that there exist no completely rigid bodies. All bodies are elastically deformable and alter in volume with changes in temperature. . . . In this connection
atomism must be borne in mind and its concept of nite divisibility. Spaces of subatomic
extension cannot be measured. Atomism compels us to give up the idea of sharply and
statically dening bounding surfaces of solid bodies. (Einstein 1952)
The starting point of quantum theory is a wave function that describes all the possible
various possible states of a particle. For example, imagine a large, irregular thundercloud
that lls up the sky. The darker the thundercloud, the greater the concentration of water
vapor and dust at that point. Thus by simply looking at a thundercloud, we can rapidly
estimate the probability of nding large concentrations of water and dust in certain parts of
the sky. The thundercloud may be compared to a single electron's wave function. Like a

222

Chapter 5
thundercloud, it lls up all space. Likewise, the greater its value at a point, the greater the
probability of nding the electron there. Similarly, wave functions can be associated with
large objects, like people. As I sit in my chair in Princeton, I know that I have a Schroedinger probability wave function. If I could somehow see my own wave function, it would
resemble a cloud very much in the shape of my body. However, some of the cloud would
spread out all over space, out to Mars and even beyond the solar system, although it would
be vanishingly small there. This means that there is a very large likelihood that I am, in
fact, sitting here in my chair and not on the planet Mars. Although part of my wave function has spread even beyond the Milky Way galaxy, there is only an innitesimal chance
that I am sitting in another galaxy. (Kaku 1995)
All our experiments indicate that quarks and bosons interact as points with no spatial
dimensions, and so are fundamental, like the leptons. If the fundamental particles are really
dimensionless points with mass, avor, color, charge, and other quantum properties, occupying no volume, then the nature of matter appears quite bizarre. The four interactions
[strong force, weak force, gravitational force, electromagnetic force] give matter shape.
Matter itself is empty. (Lederman and Schramm 1995)

Spatialization of sound particles takes two dierent approaches. First, in

synthesis and granulation, we can articulate the particle texture by scattering
individual particles in dierent spatial locations and at dierent spatial depths.
Second, through the magic of granulation, convolution, and intermodulation,
the particles can be deployed as spatializers for other sounds.
Virtual Spaces
Sound may be choreographed by positioning sources and animating their
movement. Spatialization has two facets: the virtual and the physical. In the
studio, we can spatialize sounds by imposing electronic delays, spectral lters,
panoramic motions, and reverberationgiving the illusion of sounds emerging
from imaginary environments. In the concert hall, a composer can project
sound through a multichannel or pluriphonic sound system (Roads 1998a, 2000).
These two facets are complementary.
Virtual environments can take on characteristics that would be impossible to
realize in physical architecture, such as a continuously changing echo pattern,
or the simultaneous presence of dierent qualities of ambience. Through techniques such as convolution (Roads 1992b, 1993a, 1997), we can take a sound
portrait of an existing space, such as a concert hall, and impose its spatial
characteristics onto any sound object, creating the illusion of the sound object
playing in the portrayed space. Indeed each sound object may appear to emerge

223

Transformation of Microsound

from its own virtual space. This leads to spatial counterpointthe choreography of sounds in an interplay between xed and moving positions and between
foreground and background elements. An electronic work that does not take
advantage of these possibilities may suer from a ``spatial sameness'' (Vaggione
1996a). We hear this in, for example, compositions where the voices remain
xed in space and submerged in a constant global reverberation.
Scattering Particles in Virtual Spaces
To spatialize microsound means to assign an independent spatial position to
every sonic particle. Spatial position is a function of the particle's amplitude
in two or more channels, as well the amount of reverberation in which it is
immersed. Here we concentrate on the question of position in a pluriphonic
(multichannel) environment. The next section deals with reverberation.
Two sound particles may share a spatial position, but it is also possible for
each particle in a complex sound object to occupy a unique location. This
situation creates a vivid ``three-dimensional'' sound picture, an eect that is
enhanced by loudspeakers with good imaging characteristics.
When the particle density is relatively sparse, it is possible to position each
particle manually in a sound editor or mixing program, through the manipulation of the amplitudes and panning curves of the individual tracks. But when
densities are relatively high, we design automatic scattering algorithms to assign
a position to each of thousands of particles in virtual space. These algorithms
obey high-level tendencies stipulated by the composer on a larger time scale. To
cite an example, the Cloud Generator program (described in the appendix)
oers four scattering options in a stereo eld:
1. Stationary spatial position for all grains in a cloud
2. Panoramic motion from one position to another over the duration of a cloud
3. Panoramic motion from a xed position to a random position, or vice versa
4. Random spatial position for each grain in the cloud
As pluriphonic sound diusion becomes more commonplace, we will see new
spatial scattering algorithms. With an eight-channel source, for example, an
alternative approach to spatialization is to control the density of particles per
channel. Pluriphonic sound systems that surround the audience suggest circular
and elliptical trajectories. The immersive projection spaces associated with virtual reality demand spherical scattering algorithms.

224

Chapter 5

Per-Grain Reverberation
We can deepen the spatial image by adding highly selective reverberation to
the spatial algorithm. That is, the spatialization algorithm sets the depth of
reverberation of each grain individually. This can be accomplished in several
ways.
By controlling the amplitude of a signal sent to a global reverberator, the
amount of reverberation is controlled, and so the sound's depth. Thus an ecient way to individuate the grains is to send each grain to at least two outputs.
One of the outputs is unreverberated or dry, while the other passes through a
global reverberator that is common to all grains. This is the wet signal. The
spatialization algorithm can derive the wet/dry ratio for each grain according
to a probability function. In a multichannel system, one can generalize this to N
global reverberators, where N is the number of output channels. Alberto de
Campo tested such a design at CREATE, Santa Barbara, in the octophonic
Varese Studio in the spring of 2000.
A more elaborate design is to create a bank of distinct reverberators. Some
have very short decay times, while others have long decay times. Some are dark
in color, others brighter, and so on. Probability functions determine to which
reverberator a grain is sent.
Per-grain reverberation is most striking at low densities. At high densities, the
individual reverberations fuse into a continuous background reverberation, not
much dierent from global wave reverberation.
Spatialization by Modulation with Particles
A simple way to spatialize a source sound is by modulation with particles.
Starting by generating a pattern of synthetic particles which are distributed
among two or more channels, we extract the amplitude envelope of each channel of particles then impose it on another source signal. Figure 5.5 shows the
operation applied to one channel. Here it converts a speech sound to a granular
texture. Granulation achieves a similar eect.
Convolutions with Clouds of Sonic Particles
The rolling of thunder has been attributed to echoes among the clouds; and if it is to be
considered that a cloud is a collection of particles of water . . . and therefore each capable
of reecting sound, there is no reason why very [loud] sounds should not be reverberated
. . . from a cloud. (Sir John Herschel, quoted in Tyndall 1875)

225

Transformation of Microsound

Figure 5.5 Granular modulation. (a) Granular sound. (b) Extraction of its amplitude
envelope. (c) Speech sound. (d) Speech sound modulated by the granular envelope in (b).

A class of powerful sound transformations involves convolutions of arbitrary

source sounds with clouds of sonic particles. The particles may be Gabor
grains, expodec or rexpodec grains, pulsars, trainlets, or others. The results of
such convolutions vary greatly, depending on the particles and their organization in time.
Earlier in this chapter, we showed how an accurate means of simulating the
reverberation of a given space is to convolve the impulse response of the space
with a source signal to be reverberated. A reverberator is a type of lter, where
the length in samples of the IR corresponds to the reverberation time of the
simulated hall. The IR of a room is captured by recording the room's response
to a broadband impulse. This set of samples is then convolved with the source
to be reverberated.

226

Chapter 5

In the case of convolutions with an asynchronous cloud of grains, the particles can be thought of as the IR of an unusual virtual environment (Roads
1992b). What is the shape of this environment? I imagine that it resembles a
large balloon with many long nipples. Each nipple resonates at a particular
grain frequency.
For a brief source, convolution with a sparse cloud of short grains contributes a statistical distribution of echoes. The higher the grain density of the
cloud, the more the echoes fuse into an irregular quasi-reverberation eect,
often undulating with odd peaks and valleys of intensity, and weird echoes of
the original source (gure 5.6). The virtual reection contributed by each grain
splatters the input sound in time. That is, it injects multiple delays spaced at
irregular time intervals. If each grain was a single-sample pulse, then the echoes
would be faithful copies of the original input. Since each grain may contain
hundreds of samples, however, each echo is locally ltered and time-smeared.
Time-smearing eects fall into two basic categories, depending partly on the
attack of the input sound. If the source begins with a sharp attack, each grain
generates an echo of that attack. If the cloud of grains is not continuous, these
echoes are spaced irregularly in time. If the source has a smooth attack, however, the time-splattering itself is smoothed out into a kind of strange colored
reverberation. The ``color'' of the reverberation and the echoes is determined by
the pitch and spectrum of the grains, which are a factor of the frequency, duration, envelope, and waveform of each grain (gure 5.7). See chapter 3 for
more details on grain parameters.
For low-density synchronous clouds (<10 particles/second), convolutions
result in metrical rhythms resembling tape echo, owing to the repetition of the
source. Brief particles produce clear echoes, while long particles accentuate the
bandpass ltering eect. At high densities the echoes fuse into buzzing, ringing, or rippling sonorities. The identity of the source may be obfuscated or
obliterated.
Eect of Cloud Amplitude Envelope
The amplitude envelope of a cloud plays an important role in sound transformation. If the amplitude envelope of a dense cloud of particles has an
exponential decay, then the eect is similar to granular reverberation. If the
amplitude envelope of the cloud decreases linearly or logarithmically, the reverberation sustains unnaturally. (Natural reverberation dies out quickly.) An

227

Transformation of Microsound

Figure 5.6 Reverberation by granular convolution. (a) Speech input: ``Moi, Alpha
Soixante.'' (b) Granular impulse response, consisting of one thousand 9-ms sinusoidal
grains centered at 14,000 Hz, with a bandwidth of 5000 Hz. (c) Convolution of (a) and
(b). (d) Mixture of (a) and (c) in a proportion of 5 : 1, creating reverberation around the
speech.

228

Chapter 5

Figure 5.7 Amplitude-versus-time images of convolutions with clouds of grains.

(a) Speech signal: ``It can only be attributed to human error.'' (b) Asynchronous cloud
of two hundred 10-ms grains spread across the frequency bandwidth from 60 Hz to
12000 Hz. (c) The convolution of (a) and (b) results in the speech being heard amidst
an irregular ``liquid'' echo/reverberation eect. (d) Synchronous cloud of two 10-ms
grains at 440 Hz. The circled inset shows the form of the grain in detail. (e) Convolution
of (a) and (d) results in a strongly ltered but intelligible echo of (a).

229

Transformation of Microsound

increasing amplitude envelope creates a gradual buildup in echo density over

time, unlike any in physical space.
Eect of Particle Envelope
In addition to the amplitude envelope of the entire cloud, the duration and
shape of the particle envelopes also determine the time pattern. If each particle
was only a single-sample pulse, then echoes of the source would be faithful
copies of the original. Since a particle may contain hundreds of sample points,
however, each echo is locally time-smeared according to the duration and shape
of the particle envelope.
In one experiment, I realized two dierent convolutions using identical data
sets for a dense sound cloud containing over 12,000 particles. In both cases, the
scattering algorithm was variable-density asynchronous granular synthesis.
The only dierence between the two synthetic clouds was the particle envelope.
I convolved each of these clouds with a vocal utterance: ``Lezione numero
undice, l'ora.'' In the rst case (gure 5.8a) the particle envelope is expodec, or
an instantaneous attack followed an exponential decay (Weare 1997). In the
second case (gure 5.8b) the particle envelope has a Gaussian attack and decay.
In theory, the time-smearing introduced by expodec grains should be less pronounced than that of the Gaussian grains, and this was the case. I found that
the Gaussian grains not only time-smeared, but they also introduced a churning
eect, indicated in gure 5.8b by undulations in dierent parts of the sonogram.
Eect of Particle Duration
Convolutions with short particles realize time-domain eects, while longer
grains accentuate time-smearing and ltering. For example, a stream of 1-ms
grains at 440 Hz acts mainly as an echo generator, while a stream of 10-ms
grains imposes a sharp bandpass lter. When the input sound has a smooth
attackas in a legato saxophone tonethe result is a time-varying ltering
eect on the tone. (See Roads 1993a, 1997 for more details on this technique.)
Eect of Particle Frequency
The internal frequency content of the particles imposes a strong ltering eect,
since the convolution calculates the intersection of the input spectrum and the

230

Chapter 5

Figure 5.8 Spatialization via convolution with sound particles. These sonograms are
the results of convolutions of a vocal utterance with two dense clouds of particles. The
sonograms used a 2048-point FFT with a Kaiser-Bessel window. Frequency is plotted
logarithmically from 40 Hz to 11.025 kHz. (a) The particle envelope is expodec (sharp
attack, exponential decay). (b) The particle envelope has a Gaussian attack and decay.
Notice the turgid undulations caused by time-smearing due to the smooth attack.

231

Transformation of Microsound

impulse response spectrum. If the impulse response spectrum is bandlimited, so

will be the convolved result.
Eect of Window Type in Sectioned Convolution
Many practical convolution algorithms, such as the one used in the SoundHack
program (Erbe 1995), employ sectioned convolution (described previously in
this chapter) to preserve memory space. In this technique, the convolution
applies to windowed segments of the input source. The results are equivalent to
those obtained by non-sectioned convolution, without the need to perform a
fast Fourier transform on the entire source, which could consume a great deal
of memory space. Sectioned convolution introduces another factor to the
transformation, however: the choice of window to use in sectioning the source.
Our results show that a rectangular window tends to result in a brighter and
more time-determinate convolution, while the bell-shaped Hamming window
tends to smooth both the frequency content and the time resolution.
Particle Pluriphony in Physical Spaces
The physical architecture of a concert hall and its sound system complements
the virtual acoustics of an electronic music composition. Stereophony, quadraphony, and octophony position sounds in a symmetrical lateral array in front
of or around the listener. Periphony extends this scheme to the vertical dimension (Gerzon 1973). Beginning in 1973, a new type of pluriphonic or multiloudspeaker system appeared in the form of the GMEBaphone. Designed by
Christian Clozier and Jean-Claude Le Duc, the GMEBaphone (or Cybernephone as the newest version is called) proposed an orchestra of loudspeakers
onstage, along with a complement of supporting loudspeakers surrounding the
audience. Conceived for the spatial diusion of electronic music, the system
operated under the control of a composer operating a sound-mixing console
congured for spatial projection (Clozier 1998). This idea caught on quickly in
France, and has since spread all around the world (Bayle 1989, 1993; Roads
2000).
In contrast to the typical mixdown, in which many input channels reduce to a
small number of output channels, the task of pluriphonic spatial projection is to
take in a small number of inputs and distribute them to many outputs. One can
extend the notion of pluriphony to the projection of sounds from a variety of
positions above, below, and even within the audience. The art of pluriphony
rests on four fundamental principles:

232

Chapter 5

1. The experience of an electronic music composition is greatly enhanced by a

spatial performance realized in the concert hall by the composer. The presence of a performer adds a virtuoso and dramatic aspect to the concert.
2. The sound projection system should oer a variety of contrasting spatial
images through the arrangement of multiple loudspeakers placed around the
auditorium, around the front stage, vertically, and, if possible within and
below the audience. Not all loudspeakers are in use at all times. The composer/performer makes selective use of particular spatial images to highlight
certain aspects of the work. For ensemble eects, the use of multiple loudspeakers onstage makes it possible to project a sound-image rivaling in
complexity that of an orchestra.
3. Advantage should be taken of the particularities of certain loudspeakers.
Dierent loudspeakers can integrate into the system because they each oer
a particular voicing that may be useful in articulating a specic musical
texture.
4. The illusion of stereophony, in which sound sources appear to be emerging
from between two loudspeakers, is very fragile in a large performing space.
Pluriphony seeks to replace the illusion of phantom sources with the physical
reality of many loudspeakers. Unlike illusory sources, the location of physical sources is evident from any position in the hall.
The artistic possibilities of a pluriphonic spatial performance system are most
attractive. Since 1997, we have engaged in the development of such a system at
the Center for Research in Electronic Art Technology (CREATE) at the University of California, Santa Barbara (Roads, Kuchera-Morin, and Pope 1997).
The rst version of the Creatophone (ten audio output channels, twenty loudspeakers) was inaugurated in a concert featuring Luc Ferrari in October of that
year. We have since greatly expanded the capabilities of the system.
Assessment of Spatialization on a Micro Time Scale
The real world is pluriphonic on every time scale. We enrich our music by
taking this into consideration. As a phrase unfolds, the position of each object
traces a varying topography. Spatialization imposes a curvature on an otherwise at perspective. Functional oppositions between stationary and moving
objects articulate contrapuntal relations.
In the early years of electronic music, the spatial aspect of most compositions
was xed for the entire piece. This was a macrospatial perspective. As com-

233

Transformation of Microsound

posers discovered new techniques, their spatial aesthetic became more rened.
Simultaneously, the technology of recording, editing, and mixing of sound became more sophisticated. This made it possible to associate specic localization
patterns with dierent tracks or phrases.
The digital audio workstation, introduced in the late 1980s, extended the
time scale of spatial transformations down to the level of individual sound
objects.
Through new particle scattering algorithms, micromodulation, per-grain reverberation, and convolution, we have now extended spatialization down to the
level of microsound. When we project these microspatial eects in a physical
space over widely separated loudspeakers, these tiny virtual displacements appear far larger, and the sounds dance.

Summary
Our principal metaphor for musical composition must change from one of architecture to
one of chemistry. We may imagine a new personality combing the beach of sonic possibilities, not someone who selects, rejects, classies and measures the acceptable, but a chemist
who can take any pebble, and, by numerical sorcery, separate its constituents, and merge
the constituents from two quite dierent pebbles . . . (Wishart 1994)

In the rst half of the twentieth century, Russolo (1916), Cahill (1897, 1914,
1917, 1919), Cage (1937), and Varese (1971) extended the borders of music,
allowing previously excluded sounds into the territory of composition. Mechanical noise instruments, electrodynamic tone wheels, and electronic circuits
produced these sounds. Magnetic tape recording, introduced in the 1950s, made
possible another shift in musical practice. Composers could store sounds on
tape, which opened up all the possibilities of montage.
Previously transformed sounds are another rich source for the composer.
Recursive processing, in which a transformed sound is again transformed, often
provides interesting musical evolutions. This principle of recursive variation
applies on multiple time scales. The conventional musical practice of variations
involves permutations and combinations of discrete notes: repetition, changing
the meter, changing the order of the notes, adding or omitting intervals, lling
intervals with ancillary notes, inverting the harmony, substituting chords,
transposing a melody, and so on. In contrast, sound transformation leads to
morphological changes in sound color and spatial position as well as pitch and

234

Chapter 5

rhythm. Pitch and rhythm changes introduced by recursive variation may be

continuous, rather than discrete.
By the 1980s, software made it possible to begin to develop a catalog of new
digital transformations. Automatic segmentation of sound material into microtemporal particles is the rst step in many of these techniques. Recently we
have reached a fortunate stage where many transformations operate in real
time on inexpensive computers. Besides increasing the eciency of studio work,
this means that powerful sound transformations can be performed in a virtual
or physical concert hall.
Circuit speed is less of a limiting factor, but no matter how fast computers
become, certain transformations will always be too dicult for a human being
to manipulate eectively in real time (Vaggione 1996c). Musical interfaces that
oer control through envelopes, presets, and other automation functions will
assist composers in planning detailed and elaborate transformations.

Windowed Analysis and Transformation

Overview of Windowed Spectrum Analysis

Methods of Windowed Spectrum Analysis
Models for Spectrum Analysis
Spectrum and Timbre
Data Packing and Data Reduction in Windowed Spectrum Analysis
Theory of Fourier Analysis
History of Fourier Analysis
Fourier Series
Fourier Transform
The Discrete Fourier Transform
The Short-Time Fourier Transform
Windowing the Input Signal
Operation of the STFT
Justications for Windowing
Analysis Frequencies
Time-Frequency Uncertainty
Time-Frequency Tradeos
Frequencies in between Analysis Bins
Signicance of Clutter
The Phase Vocoder
Phase Vocoder Parameters
Window Size

236

Chapter 6

FFT Size and Hop Size

Window Type
Typical PV Parameter Settings
Resynthesis Techniques
Overlap-Add Resynthesis
Oscillator Bank Resynthesis
Analysis-by-Synthesis/Overlap-Add Resynthesis
Assessment of the Phase Vocoder
Sound Transformation with Windowed Spectrum Operations
Pitch-time Changing
Frequency-Domain Filtering
Stable and Transient Extraction
Dynamic Range Manipulations
Cross-Synthesis: Vocoding, Spectral Mutation, and Analysis-Based
Formant Filtering
Other Operations on the Time-Frequency Plane
Sonographic Transformations
Sonogram Parameters
Sonogram Filtering
Assessment of Sonographic Transformations
The Tracking Phase Vocoder
Operation of the TPV
Peak Tracking
Accuracy of Resynthesis
TPV Computation and Storage
Sound Transformation with the TPV
Editing the Analysis Data
TPV Spectrum Filtering
TPV Timbre Morphing
TPV Pitch-Time Changing
Disintegration and Coalescence in the Gabor Matrix

237

Windowed Analysis and Transformation

Deterministic Plus Stochastic TPV

Experiences with TPV Transformations
Assessment of the TPV
The Vector Oscillator Transform
Assessment of the VOT
Wavelet Analysis and Resynthesis
History of Wavelet Analysis
Operation of Wavelet Analysis
Wavelet Resynthesis
Basis Functions
The Fast Wavelet Transform (FWT)
Wavelet Display
Transformation of Sounds Using Wavelets
Substituting a New Synthesis Wavelet
Wavelet Filtering
Pitch-Time Changing with the Wavelet Transform
Comb Wavelet Separation of Noise from Harmonic Spectrum
Other Wavelet Transformations
Experiences with Wavelets
Synthetic Wavelets
Assessment of the Wavelet Transform
Gabor Analysis
Theory of the Gabor Transform
Properties of the Gaussian Window Used in the GT
Musical Applications of the Gabor Transform
Assessment of the Gabor Transform
Summary
Our attention is clearly attracted by transients and movements as opposed to stationary
stimuli, which we soon ignore. Concentrating on transients is probably a strategy for
selecting important information from the overwhelming amount of data recorded by our

238

Chapter 6
senses. Yet, classical signal processing has devoted most of its eorts to the design of timeinvariant and space-invariant operators, that modify stationary signal properties. This has
led to the indisputable hegemony of the Fourier transform, but leaves aside many information-processing applications. The world of transients is considerably larger and more
complex than the garden of stationary signals. The search for an ideal Fourier-like basis
that would simplify most signal processing is therefore a hopeless quest. Instead, a multitude of dierent transforms and bases have proliferated.
Stephane Mallat (1998)

The stream of samples that form a digital audio signal is merely one representation of a microsound. To convert this signal from the time-domain to the
frequency-domain requires a stage of analysis. The analysis seeks evidence of
periodicities of a specic waveform. In the case of Fourier analysis, for example, the waveform basis is sinusoidal. Once analysis transforms the signal into a
frequency-domain representation, a large family of sonic transformations become possible. In the frequency-domain, the signal exists as a combination of
periodic functions.
What interests us here are transformations that issue from the analysis of
brief windows of sound. This chapter examines the short-time Fourier transform, the phase vocoder, the vector oscillator transform, wavelet transforms,
and the Gabor transform. Some of the explanatory material appeared in my
1996 book The Computer Music Tutorial. It appears here revised and updated.
The reports on experiments in sound transformation are new.

Overview of Windowed Spectrum Analysis

Though mathematically the Fourier theorem is beyond reproach, even experts could not at
times conceal an uneasy feeling when it came to the physical results obtained by this
method. After having for the rst time obtained the spectrum of a frequency modulated
wave, Carson (1922) wrote: ``The foregoing solutions, though unquestionably mathematically correct, are somewhat dicult to reconcile with our physical intuitions, and our
physical concepts of such `variable frequency' mechanisms, as, for example, the siren.'' The
reason is that the Fourier-integral method considers phenomena in an innite interval, sub
specie aeternitas, and this is very far from our everyday point of view. The terminology of
physics has never completely adapted itself to this rigorous mathematical denition of
``frequency.'' (Gabor 1946)

Analysis in music traditionally has referred to the study of form, phrasing, and
note relationships within a score. Digital audio technology lets us take analysis
to the level of sonic microstructureinside the note. The rst step in analysis is

239

Windowed Analysis and Transformation

windowing, or segmenting a continuous sound into discrete particles. The

window limits the duration of the analyzed signal, while imposing an envelope
on it.
Windowing is akin to synchronous granulation (see chapter 5). The use of the
term window, however, usually implies that a form of spectrum analysis will
follow. By concentrating on one grain or window of sound at a time, the analysis can focus on local or short-time spectra. Like the frames of a motion picture, a series of windowed analyses yields a portrait of the spectrum evolution
over time. The pertinence to the creative musician is that windowed operations
let us extend composition technique down to an unprecedented level of sonic
detail.
The eld of spectrum analysis has grown rapidly in the past three decades.
Analyzers have evolved from nicky instruments of acoustical study to sophisticated tools of musical production. They have emerged from institutional
laboratories into the hands of musicians working in home studios. To create
musically interesting eects, we modify the data generated by the analysis,
which results in variants of the original sound. This is analysis-resynthesis
(Risset 1989b).
Methods of Windowed Spectrum Analysis
Many spectrum analysis methods are variations on the basic technique of
Fourier analysis of component frequencies. A practical form of Fourier analysis
on digital audio signals is the short-time Fourier transform (STFT). This
method analyzes a sound by extracting successive segments (shaped by a window function) and applying a bank of lters to the selected segment. The output of each lter indicates the amplitude and the phase of the spectrum at that
particular frequency. A series of these STFTs captures a time-varying spectrum.
At the core of the STFT is the fast Fourier transform or FFT, a computationally ecient implementation of Fourier analysis (Cooley and Tukey 1965;
Rabiner and Gold 1975).
The phase vocoder (PV) (Flanagan and Golden 1966; Portno 1976, 1978,
1980, 1981; Moorer 1978; Dolson 1983, 1985, 1986) deserves special mention,
as it is a popular method of sound analysis-resynthesis. The PV converts a
sampled input signal into a time-varying spectral format. Many interesting
sound transformations can be achieved by editing and resynthesizing PV data.
Wavelet analysis is another type of windowed spectrum analysis. As opposed
to short-time Fourier analysis, with its uniform time-frequency regions, wavelet

240

Chapter 6

analysis divides the time-frequency plane into nonuniform regions. Wavelet

techniques were rst applied to audio signals in the 1980s (Grossman, et al.
1987; Kronland-Martinet, et al. 1987). Today there is a signicant body of
research on applications of wavelets to audio, with notable successes in the
domain of transient detection and data reduction.
The theories of Dennis Gabor (see chapter 2) were the precursors of windowed spectrum analysis. As we show later, Gabor analysis has been recast
recently in terms of the abstractions of wavelet theory.
Models for Spectrum Analysis
There does not seem to be any general or optimal paradigm to either analyze or synthesize
any type of sound. One has to scrutinize the soundquasi-periodic, sum of inharmonic
components, noisy, quickly or slowly evolvingand also investigate which features of the
sound are relevant to the ear. (Risset 1991)

No single method of spectrum estimation is ideal for all musical applications.

Fourier analysisthe most prevalent approachis actually a family of dierent techniques that are still evolving. A variety of non-Fourier methods continue to be developed (see Roads 1996).
Every sound analysis technique should be viewed as tting the input data to
an assumed model. Methods based on Fourier analysis model the input sound
as a sum of harmonically related sinusoidswhich it may or may not be. Other
techniques model the input signal as an excitation signal ltered by resonances,
as a sum of exponentially damped sinusoids or square waves, as a combination
of inharmonically related sinusoids, as a set of formant peaks with added noise,
or as a set of equations that represent the mechanical vibration of a traditional
instrument. Innumerable other models are conceivable. Variations in performance among the dierent methods can often be attributed to how well the
assumed model matches the process being analyzed. Hence it is important to
choose the appropriate analysis method for a particular musical application.
Spectrum and Timbre
Spectrum and timbre are related concepts, but they are not equivalent. Spectrum is a physical property that can be characterized as a distribution of energy
as a function of frequency. How to measure this energy precisely is a matter of
ongoing debate. Psychoacoustics uses the term ``timbre'' to denote perceptual

241

Windowed Analysis and Transformation

mechanisms that classify sound into families. By this denition, timbre is at

least as concerned with perception as it is with sound signals. It is certainly
easiest to discuss timbre in the realm of traditional instrument and vocal tones,
where almost all past research has focused. Only a few attempts have been
made to classify the universe of sound outside of this area, the most heroic
being the studies of Pierre Schaeer (1977; see also Schaeer, Reibel, and
Ferreyra 1967).
A common timbre groups tones played by an instrument at dierent pitches,
loudnesses, and durations. We can tell when a piano is playing, for example, no
matter what notes it plays. Human perception separates each instrument's tones
from those of other instruments played with the same pitch, loudness, and
duration. A single instrument may of course emit many timbres, as in the range
of sonorities obtained from saxophones blown in various ways.
Numerous factors inform timbre perception. These include the amplitude
envelope of a sound (especially the attack shape), undulations due to vibrato
and tremolo, formant structures, perceived loudness, duration, and the timevarying spectral envelope (Schaeer 1977; Risset 1991; McAdams and Bregman 1979; McAdams 1982; Grey 1975).
Amplitude and duration inuence the perception of timbre. A soft ute tone
at 50 dB for example, may have the same spectrum prole as one amplied to
120 dB, but we hear the latter only as a piercing blast. Similarly, a toneburst
that lasts 10 ms may have the same periodic waveshape as a tone that lasts 10
seconds, but a listener may nd it dicult to determine that they come from the
same source.
The point is that spectrum is not the only clue to perceived timbre. By
examining the time-domain waveform carefully, one can glean much about the
timbre of a sound without subjecting it to a detailed spectrum analysis.
Data Packing and Data Reduction in Windowed Spectrum Analysis
Windowed spectrum analysis, contrary to the expectations of some of its
inventors, generates an ``information explosion'' (Risset and Wessel 1982). That
is, the analysis data may occupy much more memory space than the original
input signal. Adjusting certain parameters of the analysis can make a great
dierence to the size of the analysis le. For example, a high-resolution analysis
with overlapping analysis windows greatly increases the amount of data. In
techniques such as the tracking phase vocoder (discussed later), the amount of
data depends partly on the complexity of the input sound. The analysis data

242

Chapter 6

associated with a thick noisy texture, for example, is much more voluminous
than the analysis data for a simple sinusoidal melody. Another factor in the
data explosion is the internal representation used by the analysis program,
including the word length of the numerical data.
For many reasons, there is great interest in reducing the storage requirements
of sound data. Many companies compete in the arena of digital audio encoding
schemes, which fall, broadly, into two areas: lossless packing and lossy data
reduction.
Lossless packing does not involve spectrum analysis. It makes use of redundancies in the numerical sample value to reformat it in a more memory-ecient
form. Thus it reduces storage while preserving the full integrity of the audio
data. See Craven and Gerzon (1996), or Meridian (1998) for details.
Lossy data reduction does involve windowed spectrum analysis. It dissects
sounds into a data-reduced form according to a resynthesis model, while discarding large amounts of ``nonessential'' data. In eect, it reduces sounds to a
set of control functions. It presumes the existence of a resynthesis system that
can properly interpret these control functions to reconstitute an approximation
of the sound.
Lossy data reduction schemes are built into consumer audio products such as
the Mini-Disc system, the DVD surround sound format, MP3 (MPEG 1, Layer
3) audio, and other popular Internet audio le formats. MP3, for example,
oers a variable bit rate (VBR) method (Kientzle 1998). According to the
theory of VBR, ``simple'' sound demands a low bit rate, while ``complex''
sound demands a higher bit rate. VBR encoding uses windowed spectrum
analysis and other techniques to estimate the ``complexity'' of the signal. In essence, an MP3 audio bitstream species the frequency content of a sound and
how that content varies over time. It splits the input signal into thirty-two subbands, each of which contains eighteen frequency bands, for a total of 576 frequency bands. (See Brandenburg and Bosi 1997 for details.) An MP3 player
resynthesizes the audio signal from its data-reduced form. Many points of
compromise are exploited by MP3 encoders. For example, in the interest of
speed, MP3 decoders may use integer arithmetic, which sacrices audio accuracy. The encoding of stereo information is often crude. MP3's ``joint-stereo''
mode plays the same track through both channels but with the intensity dierences of the original tracks.
Data reduction discards information. The losses may be insignicant when
the original audio program is already bandlimited, compressed in amplitude,
spatially at, distorted, and designed to be played back over a mediocre audio

243

Windowed Analysis and Transformation

system, as is the case with much popular music. But such losses are evident in
musical material that exploits the full range of a ne audio system. It is not
dicult to generate signals that reveal the weaknesses of the commercial coding
models. The degradation is more pronounced when these signals are ``copied''
(or to be more precise, recoded) or further manipulated.
For creative purposes, we prefer data reductions that leave the analysis data
in editable form. The literature of computer music includes a large body of
research work on data reduction, including pioneering studies by Risset (1966),
Freedman (1967), Beauchamp (1969, 1975), and Grey (1975). Techniques that
have been used in computer music include line-segment approximation, principal components analysis, spectral interpolation synthesis, spectral modeling
synthesis, and genetic algorithms.

Theory of Fourier Analysis

In order to comprehend many of the transformations presented later in this
chapter, it is important to have a basic understanding of the theory of Fourier
analysis. This section presents a capsule history and the essential points of the
theory.
History of Fourier Analysis
In the early eighteenth century, the French engineer and aristocrat Jean-Baptiste Joseph, Baron de Fourier (17681830), formulated a theory stating that
arbitrary periodic waveforms could be deconstructed into combinations of
simple sine waves of dierent amplitudes, frequencies, and phases. Through the
middle of the nineteenth century, Fourier analysis was a tedious task of manual
calculation. In the 1870s, the British physicist Lord Kelvin and his brother built
the rst mechanical harmonic analyzer (Marple 1987). This elaborate gearand-pulley contraption analyzed handtraced waveform segments. The analyzer
acted as a mechanical integrator, nding the area under the sine and cosine
waves for all harmonics of a fundamental period. The Michelson-Stratton harmonic analyzer (1898) was probably the most sophisticated machine of this
type. Designed around a spiral spring mechanism, it could resolve up to eighty
harmonics. It could also act as a waveform synthesizer, mechanically inverting
the analysis to reconstruct the input signal.

244

Chapter 6

In the twentieth century, mathematicians rened Fourier's method. Engineers designed analog lter banks to perform simple types of spectrum analysis.
Following the development of stored-program computers in the 1940s, programmers created the rst digital implementations of the Fourier transform
(FT), but these consumed enormous amounts of computer timea scarce
commodity in that era. Finally, in the mid-1960s, a set of algorithms known as
the fast Fourier transform or FFT, described by James Cooley at Princeton
University and John Tukey at Bell Telephone Laboratories, greatly reduced
the voluminous calculations required for Fourier analysis (Cooley and Tukey
1965).
Fourier Series
Fourier showed that a periodic function xt of period T can be represented by
the innite summation series:
xt C0

y
X
n1

Cn cosnot fn

That is, the function xt is a sum of harmonically related sinusoidal functions with the frequency on no 2p=T. C0 is the oset or DC component; it
shifts the waveform up or down. The rst sinusoidal component C1 is the fundamental; it has the same period as T. The numerical variables Cn and fn give
the magnitude and phase of each component.
A Fourier series summation is a formula for reconstructing or synthesizing a
periodic signal. But it does not tell us how to set the coecients Cn and fn for
an arbitrary input sound. For this, we need the analysis method called the
Fourier transform.
Fourier Transform
This section takes advantage of the complex exponential representation of a
sine wave at a given phase. This representation is based on these identities:
cos2p f f cos2p f j sin2p f e j2pf
So, a cosine at a given frequency and phase can also be represented as a
complex number, or a complex exponential function. (See Roads 1996, appendix A.)

245

Windowed Analysis and Transformation

Suppose that we wish to analyze a continuous-time (analog) signal xt of

innite extent and bandwidth. Fourier's theory says that xt can be accurately
reconstructed with an innite number of pure sinusoidal waves of dierent
amplitudes, frequencies, and initial phases. These waves make up the signal's
Fourier transform spectrum. The FT spectrum represents all frequencies from
0 Hz (a constant) to innity y Hz, with a mirror image in the negative
frequencies.
The formula for the FT or Fourier integral is as follows:
y
X f
xtej2pft dt
y

This says that the FT at any particular frequency f is the integral of the multiplication of the input signal xt by the pure sinusoid ej2pft . Intuitively, we
could surmise that this integral will be larger when the input signal is high in
amplitude and rich in partials. X f represents the magnitude of the Fourier
transform of the time-domain signal xt. By magnitude we mean the absolute
value of the amplitude of the frequencies in the spectrum. The capital letter X
denotes a Fourier transform, and the f within parentheses indicates that we are
now referring to a frequency-domain signal, as opposed to the time-domain
signal xt. Each value of X f is a complex number.
The magnitude is not a complete picture of the Fourier transform. It tells us
just the amount of each complex frequency that must be combined to synthesize
xt. It does not indicate the phase of each of these components. One can also
plot the phase spectrum, as it is called, but this is less often shown.
The magnitude of the Fourier transform X f is symmetric around 0 Hz.
Thus the Fourier representation combines equal amounts of positive and negative frequencies. This is the case for any real-valued input signal. This dualsided spectrum has no physical signicance. (Note that the inverse Fourier
transform takes a complex input signala spectrumand generates a realvalued waveform as its output.)
The Discrete Fourier Transform
The one kind of signal that has a discrete frequency-domain representation (i.e.,
isolated spectral lines) is a periodic signal. A periodic signal repeats at every
interval T. Such a signal has a Fourier transform containing components at a
fundamental frequency 1=T and its harmonics and its zero everywhere else.

246

Chapter 6

A periodic signal, in the precise mathematical sense, must be dened from

t y to t y. Colloquially, one speaks of signals as periodic if xt
xt T for an amount of time that is long relative to the period T. We can
construct this kind of periodic signal by replicating a nite-length signal.
Imagine that we innitely replicate the nite-length signal xt backwards and
forwards in time. In the discrete-time (sampled) domain, this produces a periodic signal xn. The use of brackets rather than parentheses indicates that the
signal is discrete, rather than continuous.
The frequency-domain representation of this replicated periodic signal xn is
called its discrete Fourier transform (DFT). The DFT provides a sampled look
at both the magnitude and phase of the spectrum of xn, and is a central tool in
musical signal processing. In eect, the DFT sets up a one-to-one correspondence between the number of input samples N and the number of frequencies
that it resolves.

The Short-Time Fourier Transform

To adapt Fourier analysis to the practical world of sampled time-varying signals, researchers molded the FT into the short-time Fourier transform or STFT
(Schroeder and Atal 1962; Flanagan 1972; Allen and Rabiner 1977; Schafer
and Rabiner 1973).
Windowing the Input Signal
As a preparation for spectrum analysis, the STFT imposes a window upon the
input signal. A window is nothing more than a simple amplitude envelope.
Windowing breaks the input signal into a series of segments that are shaped in
amplitude by the chosen window function and bounded in time according to
the length of the window function. In audio applications, the duration of the
window is usually in the range of 1 ms to 100 ms, the window envelope is bellshaped, and the segments usually overlap. By analyzing the spectrum of each
windowed segment separately, one obtains a sequence of measurements that
constitute a time-varying spectrum.
Unfortunately, windowing has the side eect of distorting the spectrum
measurement. This is because the spectrum analyzer is measuring not purely
the input signal, but rather, the product of the input signal and the window.

247

Windowed Analysis and Transformation

The spectrum that results is the convolution of the spectra of the input and the
window signals. We see the implications of this later.
Operation of the STFT
Adopting Dolson's (1986) notation, the equation for a DFT of an input signal
xm multiplied by a time-shifted window hn m is as follows:
X n; k

y
X

fxmhn mgej2p=Nkm

my

Thus the output X n; k is the Fourier transform of the windowed input at each
discrete time n for each discrete frequency band or bin k. The equation says that
m can go from minus to plus innity; this is a way of saying ``for an arbitrarylength input signal.'' For a specic short-time window, the bounds of m are set
to the appropriate length. Here, k is the index for the frequency bins, N is the
number of points in the spectrum. The following relation sets the frequency
corresponding to each bin k:
fk k=N  fs
where fs is the sampling rate. So for a sampling rate of 44.1 kHz, an analysis
window length N of 1024 samples, and a frequency bin k 1, fk is 43 Hz.
The windowed DFT representation is particularly attractive because the fast
Fourier transform or FFT can calculate it eciently.
A discrete STFT formulation indicating the hop size or time advance of each
window is:
X l; k

M
1
X

hmxm lHej2p=Nkm

m0

where M is the number of samples in the input sequence, hm is the window

that selects a block of data from the input signal xm, l is the frame index, and
H is the hop size in samples (Serra 1989).
Each block of data generated by the STFT is called a frame, by analogy to
the successive frames of a lm. Each frame contains two spectra: (1) a magnitude spectrum that depicts the amplitude of every analyzed frequency component (gure 6.1c), and (2) a phase spectrum that shows the initial phase value
for every frequency component (gure 6.1d). We can visualize each of these two
spectra as histograms with a vertical line for each frequency component along
the abscissa. The vertical line represents amplitude in the case of the magnitude

248

Chapter 6

Figure 6.1 Magnitude and phase spectra. (a) Input waveform. (b) Windowed segment.
(c) Magnitude spectrum plotted over the range 0 to 8 dB. (d) Phase spectrum plotted
over the range p to p. (After Serra 1989.)

spectrum, and the starting phase (between p and p) in the case of the phase
spectrum. The magnitude spectrum is relatively easy to read, the phase spectrum less so. When normalized to the range of p and p it is called the wrapped
phase representation. For many signals, it appears to the eye like a random
function. An unwrapped phase projection may be more meaningful visually.
(See Roads 1996, appendix A.)
To summarize, applying the STFT to a stream of input samples results in a
series of frames that make up a time-varying spectrum.
Justications for Windowing
Theory says that we can analyze a segment of any length and exactly resynthesize the segment from the analysis data. For example, we can analyze in one

249

Windowed Analysis and Transformation

pass Stravinsky's Le sacre du printemps using a thirty-minute-long window, and

reconstruct the entire piece from this analysis. This being the case, why bother
to break the analysis into overlapping windows on a micro time scale?
The reasons are several. The analysis of a thirty-minute monaural sound
sampled at 44.1 kHz would result in a spectrum of over seventy-nine million
points. A visual inspection of this enormous spectrum would eventually tell us
all the frequencies that occurred over a half hour, but would not tell us when
they occurred. This temporal information is embedded deep in the mathematical combination of the magnitude and phase spectra, hidden to the eye. Thus
the rst thing that windowing helps with is the visualization of the spectrum. By
limiting the analysis to micro segments (typically less than a twentieth of a
second), each analysis plots fewer points, and we know more accurately when
these frequencies occurred.
A second reason for using short-time windows is to conserve memory.
Breaking the input into micro segments makes it easy to calculate the FFT in a
limited memory space.
A third reason for short-time windows is that one obtains results more
quickly. For Le sacre du printemps one would have to wait up to thirty minutes
just to read in the input signal, plus however long it takes to calculate an FFT
on a seventy-nine million point input signal. Windowing the input lets one
obtain initial results quicklyafter reading just a few milliseconds of the input.
This opens up applications for real-time spectrum analysis.
Analysis Frequencies
One can think of the STFT as the application of a bank of lters at equally
spaced frequency intervals to the windowed input signal. The frequencies are
spaced at integer multiples (i.e., harmonics) of
sampling frequency
N
where N is the size of the analyzed segment. (As we will later see, the value of N
is usually greater than the actual number of sound samples analyzed; for now
we will assume they are the same length.) Thus if the sampling frequency is
50 kHz and the window length is one thousand samples, the analysis frequencies are spaced at intervals 50;000=1000 50 Hz apart, starting at 0 Hz. The
analyzer at 0 Hz measures the direct current or DC oset of the signal, a

250

Chapter 6

constant that can shift the entire signal above or below the center point of
zero amplitude.
Audio signals are bandlimited to half the sampling rate (25 kHz in this case)
and so we are concerned with only half of the analysis bins. The eective frequency resolution of an STFT is thus N=2 bins spread equally across the audio
bandwidth, starting at 0 Hz and ending at the Nyquist frequency. In our example, the number of usable audio frequency bins is ve hundred, spaced 50 Hz
apart.
Time-Frequency Uncertainty
The knowledge of the position of the particle is complementary to the knowledge of its
velocity or momentum. If we know the one with high accuracy we cannot know the other
with high accuracy. (Heisenberg 1958)

All windowed spectrum analyses are hampered by a fundamental uncertainty

principle between time resolution and frequency resolution. This is directly
analogous to a principle rst recognized by quantum physicists such as Werner
Heisenberg in the early part of the twentieth century. The linear resolution
principle (Masri, et al 1997a) states that if we want high resolution in the timedomain (i.e., we want to know precisely when an event occurs), we sacrice
frequency resolution. In other words, we can tell that an event occurred at a
precise time, but we cannot say exactly what frequencies it contained. Conversely, if we want high resolution in the frequency-domain (i.e., we want to
know the precise frequency of a component), we sacrice time resolution. This
means that we can pinpoint frequency content only over a long time interval. It
is important to grasp this fundamental relationship in order to interpret the
results of Fourier analysis.
Fourier analysis starts from this abstract premise: if a signal contains only
one frequency, then that signal must be a sinusoid that is innite in duration.
Purity of frequencyabsolute periodicityimplies innitude. As soon as one
limits the duration of this sine wave, the only way that Fourier analysis can
account for it is to consider the signal as a sum of many innite-length sinusoids
that just happen to cancel each other out in such a way as to result in a limitedduration sine wave! While this characterization of frequency neatens the
mathematics, it does not jibe with our most basic experiences with sound. As
Gabor (1946) pointed out, if the concept of frequency is used only to refer to
innitely long signals, then the concept of changing frequency is impossible.

251

Windowed Analysis and Transformation

Figure 6.2 shows the eects of time-frequency (TF) uncertainty at the juncture of an abrupt transition between two pure tones. Figure 6.2a portrays the
actual spectrum of the signal fed into the analyzer. Figure 6.2b is the measured
short-time Fourier transform of this signal. Notice the band-thickening and
blurring, which are classic symptoms of TF uncertainty.
Time-Frequency Tradeos
The FFT divides the audible frequency space into N=2 frequency bins, where N
is the length in samples of the analysis window. Hence there is a tradeo between the number of frequency bins and the length of the analysis window. For
example, if N is ve hundred and twelve samples, then the number of frequencies that can be analyzed is limited to two hundred and fty-six. Assuming
a sampling rate of 44.1 kHz, we obtain two hundred and fty-six bins equally
spaced over the bandwidth 0 Hz to the Nyquist frequency 22.05 kHz. Increasing the sampling rate only widens the measurable bandwidth, it does not
increase the frequency resolution of the analysis.
If we want high time accuracy (say 1 ms or about forty-four samples), we
must be satised with only 44/2 or twenty-two frequency bins. Dividing the
audio bandwidth from 0 to 22.05 kHz by twenty-two frequency bins, we obtain
22,050/22 or about 1000 Hz of frequency resolution. That is, if we want to
know exactly when events occur on the scale of 1 ms, then our frequency resolution is limited to the gross scale of 1000-Hz-wide frequency bands. By sacricing more time resolution, and widening the analysis interval to 30 ms, one
can spot frequencies within a 33 Hz bandwidth. For high resolution in frequency (1 Hz), one must stretch the time interval to 1 second (44,100 samples)!
Because of this limitation in windowed STFT analysis, researchers are examining hybrids of time-domain and frequency-domain analysis, multiresolution
analysis, or non-Fourier methods to try to resolve both dimensions at high
resolution.
Frequencies in between Analysis Bins
The STFT knows only about a discrete set of frequencies spaced at equal
intervals across the audio bandwidth. The spacing of these frequencies depends
on the window size. This size corresponds to the ``fundamental period'' of
the analysis. Such a model works well for sounds that are harmonic or quasiharmonic where the harmonics align closely with the bins of the analysis. What

252

Chapter 6

Figure 6.2 Time-frequency uncertainty in short-time Fourier analysis. (a) Idealized

spectrum of the signal fed to the analyzer, consisting of a tone at 1000 Hz followed immediately by a tone at 4000 Hz. (b) Analysis sonogram. The blurring indicates the zone
of time-frequency uncertainty. Both images are plotted on a logarithmic frequency scale
from 500 Hz to 7000 Hz.

253

Windowed Analysis and Transformation

happens to frequencies that fall in between the equally spaced analysis bins of
the STFT? This is the case for inharmonic sounds such as gongs or noisy
sounds such as snare drums.
Let us call the frequency to be analyzed f . When f coincides with the center
of an analysis channel, all its energy is concentrated in that channel, and so it is
accurately measured. When f is close to but not precisely coincident with the
center, energy leaks into all other analysis channels, but with a concentration
remaining close to f . The leakage spilling into all frequency bins from components inbetween bins is a well-known source of unreliability in the spectrum
estimates produced by the STFT. When more than one component is in between bins, beating eects (periodic cancellation and reinforcement) may occur
in both the frequency and amplitude traces. The result is that the analysis shows
uctuating energy in frequency components that are not physically present in
the input signal.
Signicance of Clutter
If the signal is resynthesized directly from the analysis data, the extra frequency
components and beating eects pose no problem. These eects are benign artifacts of the STFT analysis that are resolved in resynthesis. Beating eects are
merely the way that the STFT represents a time-varying spectrum in the frequency-domain. In the resynthesis, some components add constructively and
some add destructively (canceling each other out), so that the resynthesized result is a close approximation of the original.
Beating and other anomalies are harmless when the signal is directly resynthesized, but they obscure attempts to inspect the spectrum visually, or to
transform it. For this reason, the artifacts of analysis are called clutter. Dolson
(1983) and Strawn (1985) assayed the signicance of clutter in analysis of musical instrument tones. Cross-term clutter is common in higher-order analysis,
which can extract detailed phase and modulation laws embedded in the spectrum analysis (Masri, et al. 1997a, 1997b).

The Phase Vocoder

The phase vocoder (PV) uses the STFT to convert an audio signal into a complex Fourier representation. Since the STFT calculates the frequency domain
representation of the signal on a xed frequency grid, the actual frequencies of

254

Chapter 6

the partial bins have to be found by converting the relative phase change between two STFT outputs to actual frequency changes. The term ``phase'' in
phase vocoder refers to the fact that the temporal development of a sound is
contained in its phase information, while the amplitudes denote that a specic
frequency component is present in a sound. The phase contains the structural
information (Sprenger 1999). The phase relationships between the dierent bins
reconstruct time-limited events when the time-domain representation is resynthesized. The phase dierence of each bin between two successive analysis
frames determines that bin's frequency deviation from its mid frequency. This
provides information about the bin's true frequency, and makes possible a
resynthesis on a dierent time basis.
Phase Vocoder Parameters
The quality of a given PV analysis depends on the parameter settings chosen by
the user. These settings must be adjusted according to the nature of the sounds
being analyzed and the type of results that are expected. The main parameters
of the PV are:
1. Window size (also called frame size)number of input samples to be analyzed at a time.
2. FFT sizethe actual number of samples fed to the FFT algorithm; usually
the nearest power of two that is double the window size, where the unit of
FFT size is referred to by points, as in a ``1024-point FFT.''
3. Window typeselection of a window shape from among standard types.
4. Hop size or overlap factortime advance from one window onset to the
next.
Next we discuss each parameter in turn. Later we give rules of thumb for setting these parameters.
Window Size
The window size (in samples) determines one aspect of the tradeo in TF resolution. The larger the window is, the greater the number of frequency bins, but
the lower the time resolution, and vice versa. If we are trying to analyze sounds
in the lower octaves with great frequency accuracy, we cannot avoid a large
window size. Since the FFT computes the average spectrum content within a

255

Windowed Analysis and Transformation

given window, the precise onset time of any spectrum changes within the
span of the window is lost when the spectrum is plotted or transformed. (If the
signal is simply resynthesized, the temporal information is restored.) For highfrequency sounds, small windows are adequate, which are also more accurate in
time resolution.
FFT Size and Hop Size
The FFT size is typically the nearest power of two that is double the window
size. For example, a window size of 512 samples would mandate an FFT size
of 1024. The other 512 samples in the FFT are set to zeroa process called
zero-padding.
The hop size is the number of samples that the analyzer jumps along the input
waveform each time it takes a new spectrum measurement. The shorter the hop
size, the more successive windows overlap. This improves the resolution of the
analysis, but requires more computation. Some PVs specify hop size as an
overlap factor that describes how many analysis windows cover each other. An
overlap of four, for example, means that one window follows another after 25%
of the window length. Regardless of how it is specied, the hop size is usually a
fraction of the window size. A certain amount of overlap (e.g., eight times) is
necessary to ensure an accurate resynthesis. More overlap may improve accuracy when the analysis data is going to be transformed, but the computational
cost is proportionally greater.
Window Type
A spectrum analyzer measures not just the input signal but the product of the
input signal and the window envelope. The law of convolution, introduced in
chapter 5, states that multiplication in the time-domain is equivalent to convolution in the frequency-domain. Thus the analyzed spectrum is the convolution
of the spectra of the input and the window signals. In eect, the window
modulates the input signal, and this introduces sidebands or clutter into the
analyzed spectrum.
A smooth bell-shaped window minimizes the clutter. Most PVs let the user
select a window from a family of standard window types, including Hamming,
Hanning (or Hann; see Marple 1987), truncated Gaussian, Blackman-Harris,
and Kaiser (Harris 1978; Nuttall 1981). All are bell-shaped, and all work
reasonably well for general musical analysis-resynthesis. Each one is slightly

256

Chapter 6

dierent, however, so it may be worth trying dierent windows when the results
are critical. The one window to avoid is the rectangular or Dirichelet, which
introduces a great deal of clutter or extraneous frequency components into the
analyzed spectrum.
Typical PV Parameter Settings
No parameter settings of the PV are ideal for all sounds. Within a certain
range, however, a variety of traditional instrumental sounds can be analyzed
and resynthesized with reasonable delity. Here are some rules of thumb for PV
parameter settings that may serve as a starting point for more tuned analyses:
1. Window sizelarge enough to capture four periods of the lowest frequency
of interest. This is particularly important if the sound is time-stretched; too
small a window size means that individual pitch bursts are moved apart,
changing the pitch, although formants are preserved.
2. FFT sizedouble the window size, in samples.
3. Window typeany standard type except Dirichelet.
4. Hop sizeTime advance of the analysis window. If the analysis data is
going to be time-distorted, the recommended hop size is an eighth of the
frame size, in samples (i.e., eight times overlap). The minimum technical
criterion is that all windows add to a constant, that is, all data is equally
weighted. This typically implies an overlap at the 3 dB point of the particular window type chosen, from which can be derived the hop size.
Any given setting of the window size results in an analysis biased toward
harmonics of the period dened by that window size. Frequency components
that fall outside the frequency bins associated with a given window size will be
estimated incorrectly. Some analyzers try to estimate the pitch of the signal in
order to determine the optimal window size. This is called pitch-synchronous
analysis (Mathews, Miller, and David 1961). Pitch-synchronous analysis works
well if the sound to be analyzed has a basically harmonic structure.
Resynthesis Techniques
Resynthesis constructs a time-domain signal from the analysis data. If the
analysis data has not been altered, then the resynthesis should be a close simulacrum of the original signal. If the analysis data has been altered, the resyn-

257

Windowed Analysis and Transformation

thesized sound will be transformed. A variety of resynthesis techniques have

been invented. Some are more ecient, some are more accurate, some are more
robust under transformation, some are adapted for real time operation. This
section presents three techniques. The rst two appear in commonly available
phase vocoders. The third is more experimental, but gives an idea of the optimizations that can be made.
Overlap-Add Resynthesis
To resynthesize the original time-domain signal, the STFT can reconstruct each
windowed waveform segment from its spectrum components by applying the
inverse discrete Fourier transform (IDFT) to each frame. The IDFT takes each
magnitude and phase component and generates a corresponding time-domain
signal with the same envelope as the analysis window. Then by overlapping and
adding these resynthesized windows, typically at their half power or 3 dB
points, one obtains a signal that is a close approximation of the original. This is
called the overlap-add (OA) method of resynthesis.
We use the qualication ``close approximation'' as a way of comparing
practical implementations of the STFT with mathematical theory. In theory,
resynthesis from the STFT is an identity operation, replicating the input sample
by sample (Portno 1976). If it were an identity operation in practice, we could
copy signals through an STFT/IDFT any number of times with no generation
loss. However, even good implementations of the STFT lose a small amount of
information. This can be veried by a careful comparison between the input
and output waveforms. The loss may not be audible after one pass without
transformation through the STFT.
Many microsonic transformations manipulate the analysis frames before
resynthesizing the sound with the OA method. The OA process, however, is
designed for cases where the windows sum perfectly to a constant. As Allen and
Rabiner (1977) showed, any additive or multiplicative transformations that
disturb the perfect summation criterion at the nal stage of the OA cause side
eects that will probably be audible. Time expansion by stretching the distance
between windows, for example, may introduce comb lter or reverberation
eects, depending on the window size used in the analysis. Using speech or
singing as a source, some transformations result in voices with robotic or ringing artifacts. One way to lessen unwanted artifacts is to stipulate a great deal of
overlap among successive windows in the analysis stage. In selected cases, the
distortion introduced by OA resynthesis can be exploited as a sonic eect.

258

Chapter 6

Oscillator Bank Resynthesis

Oscillator bank (OB) resynthesis diers from the overlap-add approach. In
contrast to the OA model, which sums the sine waves at each frame, OB
resynthesis converts the analysis data from all analyzed frames into a set of
amplitude and frequency envelopes for multiple oscillators. In eect, the envelopes convert the analysis data from the micro time scale to the time scale of the
analyzed sound.
The advantage of OB resynthesis is that envelopes are much more robust
under musical transformation than the spectrum frames. The perfect summation criterion of the OA model does not apply in OB resynthesis. Within broad
limits, one can stretch, shrink, rescale, or shift the envelopes without worrying
about artifacts in the resynthesis process. Another strength is that the OB representation facilitates graphical editing of the spectrum. A disadvantage of OB
is that it is not as ecient computationally as OA methods.
Analysis-by-Synthesis/Overlap-Add Resynthesis
Analysis-by-synthesis/overlap-add (AS/OA) is an adaptive method designed for
improved resolution and more robust transformations. AS/OA incorporates an
error analysis procedure (George and Smith 1992). This procedure compares
the original signal with the resynthesized signal. When the error is above a
given threshold, the procedure adjusts the amplitudes, frequencies, and phases
in the analysis frame to approximate the original more closely. This adaptive
process can occur repeatedly until the signal is more-or-less precisely reconstructed. As a result, the AS/OA method can handle attack transients, inharmonic spectra, and eects such as vibrato with greater accuracy than the OA
method. It also permits more robust musical transformations. Another method
called the tracking phase vocoder, presented later in the chapter, has similar
benets.
Assessment of the Phase Vocoder
Present methods of windowed spectrum analysis owe a debt to Dennis Gabor's
pioneering work (1946, 1947, 1952; see also chapter 2). The frames of the STFT
are analogous to his acoustical quanta. The projection of the time-frequency
plane onto the sonogram is analogous to a visual representation of the Gabor
matrix.

259

Windowed Analysis and Transformation

The phase vocoder has emerged from the laboratory to become a popular
tool. It is packaged in a variety of widely distributed musical software. The
compositional interest of the PV lies in transforming the analysis data before
resynthesis, producing variations of the original sound. What the composer
seeks in the output is not a clone of the input, but a musical transformation that
maintains a sense of the source's identity.
The weaknesses of the STFT and the PV as representations for sound
are well known. The uncertainty principle pointed out by Gabor is embedded
deeply within the STFT. Time-frequency information is smeared. Overlapping
windows mean that it is impossible to modify a single time-frequency atom
without aecting adjacent atoms. Such a change will most likely lead to a discontinuity in the resynthesized signal. Many transformations sound ``blurry'' or
``sinusoidal'' in quality, a common artifact of Fourier techniques in general.
The tracking phase vocoder, described later, is a more secure transformation
tool, but it has its own imperfections.
On the positive side, the PV is powerful. Good implementations of the PV
oer the possibility of modifying pitch, time, and timbre independently.

Sound Transformation with Windowed Spectrum Operations

Inside a windowed spectrum analyzer is a granulated time-frequency representation of sound. By manipulating this representation, we can obtain many
transformations that are dicult or impossible to achieve with time-domain
procedures, including high-quality pitch-time changing, frequency-domain
ltering, stable and transient extraction, multiband dynamics processing, crosssynthesis, and many other exotic eects. This section explores these techniques,
many of which are implemented within phase vocoders.
Pitch-Time Changing
One of the most common windowed spectrum transformations is the altering of
a signal is duration whilst maintaining its original pitch. Inversely, one can
change pitch without altering duration. For these eects, the phase vocoder
often achieves better sound quality than can be obtained with the time-domain
granulation algorithms described in chapter 5.
In a PV with overlap-add resynthesis, the time stretching/shrinking algorithm
moves the onset times of the overlapping frames farther apart (when stretching)

260

Chapter 6

or closer together (when shrinking) in the resynthesis. For the smoothest transpositions, the PV should multiply the phase values by the same constant used in
the time base changing (Arb 1991).
Pitch-shifting alters the pitch without changing the time base. Pitch-transposition is a matter of scaling the frequencies of the resynthesis components. For
speech signals in particular, however, a constant scale factor changes not only
the pitch but also the formant frequencies. For upward shifts of an octave or
more, this reduces the speech's intelligibility. Thus Dolson (1986) suggested a
correction to the frequency scaling that reimposes the original spectral envelope
on the transposed frequency spectrum. If the original spectrum had a formant
at 2 kHz, for example, then so will the transposed version.
Frequency-Domain Filtering
Spectrum lters operate in the frequency domain by rescaling the amplitudes of
selected frequency bins. Some of their controls are similar to traditional timedomain lters, such as center frequency, bandwidth, and gain or boost in dB.
Other controls apply to the windowed analysis, such as the window size, FFT
size, and overlap factor. These controls aect the eciency and quality of the
analysis-resynthesis process. For example, longer windows and FFTs generally
result in more pronounced ltering eects.
Dierences between time-domain lters and spectral lters show up when
the bandwidths are narrow and the lter Q is high. The spectral lter breaks
down a broadband signal into individual sinusoidal components. A tell-tale
``breebles'' artefact may be heard, as individual components pop in and out.
Breebles are characteristic of manipulations on windowed Fourier analyses in
general, and appear in a number of other PV transformations.
Another approach to frequency-domain ltering provides a graphic interface,
in which the user sees a sonogram display of the sound. The software provides a
palette of drawing tools that let users erase or highlight selected regions of the
sonogram image. The sound is then resynthesized on the basis of the altered
sonogram image. (See the later section on sonographic transformations.)
Stable and Transient Extraction
This is a class of transformations that sorts audio waveforms on a micro time
scale into two categories: stable and transient. Spoken vowels, for example, are
relatively stable frequencies compared to the transient frequencies in conso-

261

Windowed Analysis and Transformation

nants. Once these frequencies are separated, the signals can be further manipulated individually.
In Erbe's (1995) implementation of stable and transient extraction, the user
species:
1. Number of bands in the analyzer
2. Number of frames to analyze at a time
3. Frequency threshold for the transient part of the signal
4. Frequency threshold of the stable part of the signal
For example, the transient part could be specied as changing more than 30
Hz per frame, while the stable part is specied as changing less than 5 Hz per
frame. Erbe's extraction algorithm takes the average of the change in instantaneous frequency over several FFT frames. If this average is greater than the
stipulated value for transient information, the amplitude and phase from the
source is assigned to the transient spectrum. Similarly, if the average change is
less than the stipulated value for stable information the amplitude and phase
from the source is assigned to the stable spectrum. Note that if the transient and
stable thresholds are not identical, this leaves behind a part of the spectrum that
is between the twoneither stable nor transient.
Another approach to stable and transient extraction is via spectral tracing
(Wishart 1994; Norris 1997). Spectral tracing analyzes a sound and retains only
the loudest or softest N% of the partials in the spectrum. To extract the transient part of a spoken voice, one retains only the softest 1% of the analyzed
spectra. The sound quality of this 1% after the result is high-pass ltered, is like
noisy whispering.
Dynamic Range Manipulations
Spectrum analysis makes it possible to manipulate the dynamic range of
selected frequency bands. The reader is referred to the discussion of dynamics
processing on a micro time scale in chapter 5.
Cross-Synthesis: Vocoding, Spectral Mutation, and Analysis-Based Formant
Filtering
Cross-synthesis extracts characteristics from the spectrum of one signal and uses
them to modify the spectrum of another signal. This can take a variety of forms,
including, among others, vocoding, spectral mutation, and formant ltering.

262

Chapter 6

Vocoding (as opposed to phase vocoding) analyzes one signal to adjust a

bank of lters (or subbands) that are applied to a second signal. Its operation is
relatively simple, meaning that it can be computed in real time. First, it extracts
the amplitude envelope of the input signal coming through each lter. This is
typically accomplished by rectication and lowpass ltering. The signal from
which the amplitude is extracted is called the modulator; the signal that is
aected is called the carrier. It then applies the amplitude envelopes from the
modulator to a second lter bank through which the carrier is passing. In a
typical vocoder, there are a relatively small number of lters (<50) and the
center frequency of each of the lters is xed. These constraints give the eect
its characteristic sound quality.
An example of a vocoder is found in Arboretum's Hyperprism program
(MacOS). This 26-band real-time vocoder performs an FFT on successive windows of the modulator signal. It adjusts the amplitude of its lter bands to
match the spectrum of the incoming signal. That is, the modulating signal sets
the gain for each of the lters. It then applies the 26-band lter bank to the
carrier signal.
Another type of cross-synthesis involves mutations from one set of spectral
data to another. The spectral mutation operations in the SoundHack program
(MacOS) interpolate the sign or magnitude of a source set of spectral frames
into the sign or magnitude of a target set (Polansky and Erbe 1996). The mutation functions operate on the phase and amplitude data of each analyzed
frequency band. A single frame spans a microtemporal duration that is usually
between 10 and 90 milliseconds. The degree of mutation is called W, which
takes a value between 0 (source) and 1 (target). The software oers ve mutation functions, some of which move linearly from one the source to the target,
while others scramble the two.
A third possibility is creating a new spectrum by taking the intersection of
two source spectra. This can be achieved in a variety of ways. One of the spectra, for example, might be derived from an analysis based on linear predictive
coding or LPC (Roads 1996). LPC analysis calculates an overall spectrum
envelope. One could use the LPC spectrum of one sound to shape the detailed
spectrum of another sound, where the detailed analysis was derived by a phase
vocoder (Serra 1997). U&I Software's MetaSynth program (MacOS) takes
another approach to this cross-synthesis (Wenger and Spiegel 1999). Its Formants Filter eect uses a 128-band lter bank to nd the formant peaks in the
spectrum of one signal. It then applies this bank to boost the same formant
regions in another signal. The eect thus emphasizes the frequencies that the
two sounds have in common.

263

Windowed Analysis and Transformation

Other Operations on the Time-Frequency Plane

Over the past decade, an extensive catalog of transformations based on windowed spectrum analysis has been developed. Many of these were originally
implemented in the Composer's Desktop Project (CDP) software package for
Windows computers (Endrich 2000). Trevor Wishart's book Audible Design
(1994) describes many of these eects using evocative drawings. Table 6.1 is a
list of spectrum-based operations available in the CDP package, grouped into
fourteen categories. Some of the operations are utilitarian, and do not transform the sound.
For those working on MacOS computers, at least a dozen software packages
provide spectrum analysis and resynthesis capabilities. Many of the CDP
spectrum operations, for example, were made available as plugins by Michael
Norris (1997) and Alex Yermakov (1999) within the SoundMaker program
(Ricci 1997).

Sonographic Transformations
A sonogram, sonograph, or spectrogram is a well-known spectrum display technique. It represents a sound signal as a two-dimensional display of time versus
``frequency amplitude.'' That is, the vertical dimension depicts frequency
(higher frequencies are higher up in the diagram) and shades of gray (or color)
indicate the amplitude within a given frequency band.
The rst sonogram was Backhaus's (1932) system. In the 1950s, the Kay
Sonograph was a standard device for printed sonograms which combined a
bank of narrow bandpass analog lters with a recording system that printed
dark traces on a roll of paper. The bars grew thicker in proportion to the energy
output from each lter.
Today, the STFT is at the core of sonogram analysis. The sonogram can be
applied on various time scales (gure 6.3). A sonogram of the meso time level
(gure 6.3a) portrays general features such the onset of notes or phonemes,
their pitch, formant peaks, and major transitions. See Cogan (1984) for an
example of using sonograms in the analysis of musical mesostructure. A sonogram of a single note (gure 6.3b) reveals its pitch and spectrum. A sonogram
of a single particle (gure 6.3c) is a coarse indicator of spectrum, since frequency resolution is poor on a micro time scale.

264

Chapter 6
Table 6.1 Typical spectrum operations in the Composer's Desktop Project software
1. Blur analysis data
Blur avrgAverage spectral energy over N adjacent channels
Blur blurBlur spectral data over time
Blur chorusAdd random variation to amplitude or frequency in analysis channels
Blur drunkModify sound by a drunken walk (a probabilistic process) along analysis
windows
Blur noiseAdd noise to spectrum
Blur scatterThin the spectrum in a random fashion
Blur shueShue analysis windows according to a specic scheme
Blur spreadSpread spectral peaks
Blur suppressSuppress the most prominent channel data
Blur weaveWeave among the analysis windows in a specied pattern
2. Combine analysis data from two or more les
Combine crossReplace spectral amplitudes of rst le with those of second
Combine diFind (and retain) the dierence between two spectra
Combine leafInterleave (groups of ) windows of several spectra
Combine makeGenerate an analysis le from data in a formant data le and a pitch
data le
Combine maxRetain loudest channel components per window amongst several
spectra
Combine meanGenerate the mean of two spectra
Combine sumAdd one spectrum to another
3. Focus on features of analysis data
Focus accuSustain each spectral band, until louder data appears in that band
Focus exagExaggerate the spectral contour
Focus focusFocus spectral energy onto the peaks in the spectrum
Focus foldOctave-transpose spectral components into a specied frequency range
Focus freezeFreeze the spectral characteristics in a sound, at given times, for
specied durations
Focus stepStep-frame through a sound by freezing the spectrum at regular time
intervals
4. Formant operations
Formants getExtract evolving formant envelope from an analysis le
Formants getseeGet formant data from an analysis le and write as a le for viewing
Formants putImpose formants in a formant data le on the spectrum in an analysis
le
Formants seeConvert formant data in binary formant data le to a le for viewing
Formants vocodeImpose spectral envelope of one sound onto another sound

265

Windowed Analysis and Transformation

Table 6.1 (continued)
5. Highlight features of analysis data
Hilite arpegArpeggiate the spectrum
Hilite bandSplit spectrum into bands and process these individually
Hilite bltrTime-average and trace the spectrum
Hilite lterHighpass, lowpass, bandpass, and notch lters, on spectral data
Hilite greqGraphic equalizer lter on the spectrum
Hilite pluckEmphasize spectral changes (e.g., use with Hilite arpeg)
Hilite traceHighlight N loudest partials, at each moment (window) in time
6. Morph smooth transitions between sounds
Morph bridgeMake a bridging interpolation between two sound spectra by
interpolating between two time-specied windows in the two input les
Morph glideInterpolate linearly between two single analysis windows extracted with
Spec grab
Morph morphMorph between one spectrum and another, where the spectra may be
time-varying
7. Pitch operations
Pitch altDelete alternate harmonics
Pitch chordSuperimpose transposed versions of the spectrum within the existing
spectral envelope
Pitch chord2Superimpose transposed versions of a sound on the original sound
Pitch octOctave transpose without a formant shift
Pitch pickRetain channels that might hold specied partials
Pitch shiftShift pitch of (part of ) the spectrum
Pitch tuneReplace spectral frequencies by harmonics of specied pitch(es)
8. Retrieve pitch information
Pitchinfo infoDisplay information about pitch data in pitchle
Pitchinfo hearConvert binary pitchle to analysis test tone le (resynthesize to hear
pitch)
Pitchinfo seeConvert binary pitchle or transposition le to a pseudo-soundle, for
viewing
Pitchinfo writeConvert a binary pitch data le to a time frequency breakpoint text
le
Pitchinfo zerosDisplay whether a pitch le contains uninterpolated zeros (unpitched
windows)
9. Phase vocoder analysis and resynthesis of sounds
Pvoc analConvert soundle to spectral le
Pvoc extractAnalyze, then resynthesize sound with various options
Pvoc synthConvert spectral le to soundle

266

Chapter 6
Table 6.1 (continued)
10. Repitch pitch-related data
Repitch approxMake an approximate copy of a pitchle
Repitch combineGenerate transposition data from two sets of pitch data, or
transpose pitch data with transposition data, or combine two sets of transposition data
to form new transposition data, producing a binary pitch data le output
Repitch combinebGenerate transposition data from two sets of pitch data, or
transpose pitch data with transposition data, or combine two sets of transposition data
to form new transposition data, producing a time value breakpoint le output
Repitch cutCut out and keep a segment of a binary pitch data le
Repitch exagExaggerate pitch contour
Repitch getpitchExtract pitch from spectrum to a pitch data le
Repitch imposeTranspose spectrum (spectral envelope also moves)
Repitch imposefTranspose spectrum, but retain original spectral envelope
Repitch invertInvert pitch contour of a pitch data le
Repitch quantizeQuantize pitches in a pitch data le
Repitch randomizeRandomize pitch line
Repitch smoothSmooth pitch contour in a pitch data le
Repitch transposeTranspose pitches in a pitch data le by a constant number of
semitones
Repitch vibratoAdd vibrato to pitch in a pitch data le
11. Spectrum operations
Spec bareZero the data in channels that do not contain harmonics
Spec cleanRemove noise from phase vocoder analysis le
Spec cutCut a section out of an analysis le, between starttime and endtime
(seconds)
Spec gainAmplify or attenuate the spectrum
Spec grabGrab a single analysis window at time point specied
Spec limitEliminate channel data below a threshhold amplitude
Spec magnifyMagnify (in duration) a single analysis window and time time to
duration dur
12. Retrieve spectrum information
Specinfo channelReturns phase vocoder channel number corresponding to specied
frequency
Specinfo frequencyReturns center frequency of phase vocoder channel specied
Specinfo levelConvert (varying) level of analysis le to a le, for viewing
Specinfo octvuText display of time varying amplitude of spectrum, within octave
bands
Specinfo peakLocate time varying energy center of spectrum (text display)
Specinfo printPrint data in an analysis le as text to le
Specinfo reportText report on location of frequency peaks in the evolving spectrum
Specinfo windowcntReturns the number of analysis windows in input le

267

Windowed Analysis and Transformation

Table 6.1 (continued)
13. Strange operations
Strange glisCreate glissandi inside the (changing) spectral envelope of the original
sound
Strange invertInvert the spectrum
Strange shiftLinear frequency shift of (part of ) the spectrum
Strange waverOscillate between harmonic and inharmonic state
14. Stretch time or frequency data
Stretch spectrumStretch or compress the frequencies in the spectrum
Stretch timeStretch or compress a sound in time without changing the pitch

Sonogram Parameters
The parameters of the modern sonogram are the same as those of the STFT,
except for the display parameters. Adjustments to these parameters make a
great dierence in the output image:
1. Range of amplitudes and the type of scale used, whether linear or
logarithmic.
2. Range of frequencies and the type of scale used, whether linear or
logarithmic.
3. Window size (number of samples to analyze) and the size of the FFT; the
resolution of time and frequency depend on these parameters.
4. Time advance of the analysis window (hop size) in samples or window
overlap factor. This determines the time distance between successive columns in the output display.
5. Number of frequency channels to display, which determines the number of
rows in the graphical output and is related to the range and scale of the frequency domain; this cannot exceed the resolution imposed by the window
size.
6. Window typesee the previous discussion in the section on the phase
vocoder.
The window parameters (3) have the most dramatic eect on the display. A
short window results in a vertically oriented display, indicating the precise onset
time of events but blurring the frequency reading. A medium length window
resolves both time and frequency features fairly well, indicating the presence of

268

Chapter 6

Figure 6.3 Sonograms on three time scales. (a) Meso time scale. 13-second sonogram of
a J. S. Bach suite transcribed for recorder, played in a reverberant environment. The
frequency scale is linear to 4 kHz. (b) Sound object time scale. 300-ms tone played on an
Ondioline, a vacuum tube electronic music instrument (Jenny 1958). The frequency scale
is linear to 5 kHz. (c) Micro time scale. 50-ms view on a pulsar particle. The frequency
scale is linear to 12 kHz.

269

Windowed Analysis and Transformation

formant frequencies. A long window generates a horizontally oriented display,

as individual frequency bands come into clear view, but their position in time
smears along the horizontal axis.
Sonogram Filtering
A sonogram is a projection of the Gabor time-frequency matrix. So to manipulate the sonogram image is to manipulate the sound. We looked at sonographic synthesis in chapter 4. To draw on the sonogram is to inscribe sound
directly, correspondingly, to erase part of a sonogram is to lter the sound. A
pioneering program with this capability was Gerhard Eckel's SpecDraw (Eckel
1990).
Today a variety of programs allow for manipulation of the sonogram image.
For example, the Ircam AudioSculpt (MacOS) application (Serra 1997) lets
users extract or erase spectra with graphical drawing tools. One can inscribe an
arbitrary polygon on the time-frequency plane, and stipulate a boost or attenuation of the selected frequencies in that region (see gure 8.1). Users can also
boost or erase with a pen tool.
Applications like U&I Software's MetaSynth provide a palette of brushes
and graphical operations for manipulating the sonogram image (table 6.2).
Going further, one can import and export sonogram images in standard
graphics le formats, and then alter them with image processing software, such
as Adobe PhotoShop with its many plugin eects. This opens up vast possibilities of sonographical treatments.
Assessment of Sonographic Transformations
Sonograms trace a sound's morphologyits internal shape in frequency and
time, making them powerful musical resources in sound transformation.
Moreover, a sonogram is an intuitive musical interface. It is easy to learn to
draw and erase a sonogram image. This apparent ease of use can be deceptive,
however; as with any other musical tool, it takes practice to achieve good
results with sonographic transformations. A pretty image does not necessarily
correspond to an interesting sound. A sonogram is the projection of a mathematical transform and does not correspond exactly to what we hear. What we
perceive as a fused sound object often shows up as a splintered branch of
separate frequency components. The ear is quite sensitive to details within this
branch, including the frequency ratios and the relative strength of each component. A similar drawing would not necessarily produce a similar sound.

270

Chapter 6
Table 6.2 Sonographic transformations in MetaSynth
Sonographical brushes (the sizes of the brushes are variable)
PenHard-edged rectangular brush
Air brushRound-edge brush with translucent edges
Filter brushThis brush acts as a multiplier, brightening or darkening depending on
the selected color
Harmonics brushPaints a fundamental and up to ve harmonics
Attack brushPaints a sharp left edge and then a soft decay
Smoothing brushSmooths the pixels over which it passes
Spray brushSprays a cloud of grains
Decay brushExtends existing pixels to the right to elongate their decay
Note brushLeaves a trail of discrete notes, aligned to the brush grid interval
Line brushPaints a harmonic line across the width of the canvas
Smear brushSmears existing pixels
Smear brighter brushSmears existing pixels with a brighter gradient than the smear
brush
Clone brushCaptures pixels under the brush when the mouse button is pressed and
then paints with the captured pixels
Sonographical transformations on the time-frequency grid
Cut, copy, paste selected part of sonogram image
X-Y scalingTime stretch or compress, frequency stretch or compress
Shift up or downTranspose all frequencies
RotateChange the direction of all frequency trajectories
Contrast and luminenceRemoves low-level frequencies, or amplies noise
Octave transposeShift the image (and sound) down by an octave
SmoothBlur the image so that the sound becomes more sinusoidal
InvertWhite becomes black, etc.; a sound becomes a silence in a noise eld
Max pictPaste the PICT clipboard into the selected area, treating the clipboard's
black pixels as transparent. Where there are coincident pixels, the brightest one is kept.
(Note: PICT is a MacOS bitmap graphics format.)
Min pictPaste the PICT clipboard into the selected area. Where there are coincident
pixels, the one of lowest amplitude one is kept. If either image has black, the result is
black.
Add pictCombine the luminosities of two images
Subtract pictSubtract the contents of the PICT clipboard from the selected area
Multiply pictMultiple the selected region by the contents of the PICT clipboard
Merge pictMerge the PICT clipboard with the selected region using a 50% blend
Crossfade pictCrossfade the PICT clipboard with the selected region left to right
ExpandVertically expand the pixel spacing
Fade in out pictFade the PICT clipboard in then out while also fading the selected
region out then in

271

Windowed Analysis and Transformation

Table 6.2 (continued)
ContractVertically contract the region between the pixels
Triangle lterA ``solarization'' eect where luminence above 50% is changed to black
Emboss left or rightEmphasize the leading or trailing edges of shapes
Trace edgesSharpen the image, making the time-frequency structure more clear
Sonographical transformations on the time grid
mBlurSmooth transients, remove short pixels
QuantizeForce notes to start and end at the positions determined by the current grid
interval
EchoRepeat pixels to the right with decreasing brightness
Pre-echoRepeat pixels to the left with decreasing brightness
ReverbSimulate reverberation by extending and fading the duration of all pixels in
the selected region
RepeatRepeat pixels identically across the region according to the stipulated grid
ReverseReverse the region
PulseRemove pixels according to the stipulated grid interval
SawSimilar to pulse, but fades out rather than cuts
ShortenShorten the duration of the pixel lines
AttacksSharpen the attacks of all notes

Furthermore the tools provided with sonographic interfaces are not always
precise, making it dicult to achieve predictable results. To edit on a micro
level, we may want to zoom in to the sonogram image. When we do so, however, the image pixellates into large blocks. This simply reects the hard fact
that a sonogram is intrinsically limited in time-frequency resolution.

The Tracking Phase Vocoder

Certain implementations of the PV are called tracking phase vocoders (TPVs)
because they follow or track the most prominent peaks in the spectrum over
time (Dolson 1983; McAulay and Quatieri 1986; Serra 1989; Maher and
Beauchamp 1990; Walker and Fitz 1992). Unlike the ordinary phase vocoder,
in which resynthesis frequencies are limited to harmonics of the analysis
window, the TPV follows changes in frequencies. The result of peak tracking
is a set of amplitude and frequency envelopes that drive a bank of sinusoidal
oscillators in the resynthesis stage.
The tracking process follows only the most prominent frequency components. For these components, the result is a more accurate analysis than that

272

Chapter 6

done with an equally spaced bank of lters (the traditional STFT implementation). Another benet is that the tracking process creates frequency and
amplitude envelopes for these components, making them more robust under
transformation than overlap-add frames. A disadvantage is that the quality of
the analysis depends more heavily on proper parameter settings than in the
regular STFT. It may take multiple attempts to tune the analysis parameters for
a given sound.
Operation of the TPV
A TPV carries out the following steps:
1. Compute the STFT using the frame size, window type, FFT size, and hop
size specied by the user.
2. Derive the squared magnitude spectrum in dB.
3. Find the bin numbers of the peaks in the spectrum.
4. Calculate the magnitude and phase of each frequency peak.
5. Assign each peak to a frequency track by matching the peaks of the previous
frame with those of the current frame (see the description of peak tracking
later).
6. Apply any desired modications to the analysis parameters.
7. If additive resynthesis is requested, generate a sine wave for each frequency
track and sum all sine wave components to create an output signal; the
instantaneous amplitude, phase, and frequency of each sinusoidal component is calculated by interpolating values from frame to frame (or use the
alternative resynthesis methods described earlier).
Peak Tracking
The tracking phase vocoder follows the most prominent frequency trajectories
in the spectrum. Like other aspects of sound analysis, the precise method of
peak tracking should vary depending on the sound. The tracking algorithm
works best when it is tuned to the type of sound being analyzedspeech, harmonic spectrum, smooth inharmonic spectrum, noisy, etc. This section briey
explains more about the tracking process as a guide to setting the analysis
parameters.
The rst stage in peak tracking is peak identication. A simple control that
sets the minimum peak height focuses the identication process on the most

273

Windowed Analysis and Transformation

signicant landmarks in the spectrum. The rest of the algorithm tries to apply a
set of frequency guides which advance in time. The guides are only hypotheses;
later the algorithm will decide which guides are conrmed frequency tracks.
The algorithm continues the guides by nding the peak closest in frequency to
its current value. The alternatives are as follows:
1 If it nds a match, the guide continues.
1 If a guide cannot be continued during a frame it is considered to be
``sleeping.''
1 If the guide does not wake up after a certain number of frameswhich may
be specied by the userthen the tracker deletes it. It may be possible
to switch on guide hysteresis, which continues tracking a guide that falls
slightly below a specied amplitude range. Guide hysteresis alleviates the
audible problem of ``switching'' guides that repeatedly fade slightly, are
cut to zero by the peak tracker, and fade in again (Walker and Fitz 1992).
With hysteresis the guide is synthesized at its actual value instead of at zero
amplitude.
1 If there is a near match between guides, the closest wins and the ``loser''
looks for another peak within the maximum peak deviation, a frequency band
specied by the user.
1 If there are peaks not accounted for by current guides, then a new guide
begins.
Windowing may compromise the accuracy of the tracking, particularly in
rapidly moving waveforms such as attack transients. Processing sounds with a
sharp attack in time-reversed order helps the tracking algorithm (Serra 1989).
This gives the partial trackers a chance to lock onto their stable frequency trajectories before meeting the chaos of the attack, which results in less distortion.
The data can be reversed back to its original order before resynthesis.
Accuracy of Resynthesis
In contrast to the myth of ``perfect reconstruction'' which pervades the mathematical theory of signal processing, the actual quality of all analysis-resynthesis
methods is limited by the resolution of the input signal and the numerical
precision of the analysis procedures. Distortions are introduced by numerical
roundo, windowing, peak-tracking, undersampling of envelope functions, and
other aspects of the analysis. Compact disc quality audio (16-bit samples, 44.1

274

Chapter 6

kHz sampling rate) poses a priori limits on the frequency response and dynamic
range of the analysis. The fast Fourier transform on a 16-bit signal has a limited
dynamic range, and the TPV reduces this further by discarding spectral information below a certain threshold. Any modication that changes the frequency
characteristics will likely result in aliasing. Any reduction in amplitude caused
by enveloping reduces the bit resolution. The resynthesized result may have lost
a large portion of its dynamic range, and artefacts such as amplitude gating,
distortion, aliasing, and graininess are not uncommon.
In a well-implemented nontracking phase vocoder, when the analysis
parameters are properly adjusted by a skilled engineer and no modications are
made to the analysis data, the error is perceptually negligible. The TPV, on the
other hand, discards information that does not contribute to a track. If the TPV
parameters are not adjusted correctly, the sifting of low-level energy may discard signicant portions of the sound, particularly noisy, transient energy. This
can be demonstrated by subtracting an analysis of the resynthesized signal from
an analysis of the original signal to yield a residual (Serra 1989). One can consider this residual to be analysis/resynthesis error. It is common to refer to the
resynthesized, quasi-harmonic portion as the ``clean'' part of the signal and the
error or noise component as the ``dirty'' part of the signal. For many sounds
(i.e., those with fast transients such as cymbal crashes), the errors are quite
audible. That is, the clean signal sounds unnaturally sanitized or sinusoidal,
and the dirty signal, when heard separately, contains the missing grit.
TPV Computation and Storage
TPV analysis consumes large quantities of computer power even though the
inner core is implemented using the ecient FFT algorithm. It can also generate a large amount of analysis data, often ten times greater than the size of the
sample data being analyzed. The size of the analysis depends on many factors,
however. Low-level narrowband sounds require fewer tracks. Dierent settings
of the analysis parameters can greatly aect the analysis le size.
Sound Transformation with the TPV
The TPV's representation of sounda bank of hundreds of oscillators driven
by amplitude and frequency envelopesis a robust one that lends itself to
many evocative transformations.

275

Windowed Analysis and Transformation

Editing the Analysis Data

Certain TPV implementations let one directly edit the analysis datathe frequency, amplitude, and phase curves (Moorer 1978; Dolson 1983). This laborious process is aided by automatic data reduction procedures and graphical
editors. For example, Lemur Pro (MacOS) provides a two-dimensional (timefrequency) spectrum editor that lets users modify the characteristics of individual frequency tracks or groups of tracks (Walker and Fitz 1992; Fitz, et al.
1995).
TPV Spectrum Filtering
When the TPV is coupled with a graphical interface for zooming in and selecting tracks, one can alter the amplitude of specic tracks. It may also be possible
to cut or extract certain tracks. Cutting removes them from the spectrum, while
extracting removes everything but the selected tracks. Such alterations have the
eect of ltering the spectrum.
A single sound object may contain hundreds of closely-spaced tracks (gure
6.4). So the question becomes how to select tracks to lter. Keep in mind that
an operation on an individual track makes little dierence, it is only when a
band of frequencies are operated on that the eect becomes noticeable.
Selecting tracks manually is a labor-intensive way to lter a sound. The key
to eciency in TPV editing is to use other strategies.
Selecting tracks by their length, for example, leads to potentially interesting
transformations on a micro scale. One interesting transformation is based on
eliminating all tracks greater than a threshold of, say, 70 ms, leaving only the
microtemporal tracks. This has the eect of removing pitch information, so that
a speaking voice becomes wheezing and whispering. Selecting and boosting all
tracks less than 10 ms produces a gravely sounding voice.
Another way to select tracks is by amplitude, so that only tracks over or
under a given threshold are selected. In combination with the length criterion,
this can be used to granulate a sound le in time and frequency. Note that the
tracking algorithm in TPV analysis is already a lter. By tuning the parameters
so that the analysis rejects low-amplitude frequency components, for example,
the analysis data is strongly ltered. One imagines that more sophisticated track
selection principles will become available in the future, such as selecting harmonically related tracks.

276

Chapter 6

Figure 6.4 This gure shows the profusion of tracks associated with even a single
sound, in this case, the French word payons (we pay).

TPV Timbre Morphing

One type of cross-synthesis is called timbre morphing or just morphing. The term
``morphing,'' like the term ``cross-synthesis,'' is not a precise one. It describes
a variety of processes that combine the properties of two sounds. Morphing
combines the characteristics of two analyzed les to create a new le with an
intermediate timbre. For instance, if a long loud tone with a fast and narrow
vibrato is morphed with a short quiet tone with a slow and wide vibrato, the
ideal morphed result might be a medium length, medium loudness tone with an
intermediate vibrato speed and width. (But as the phrase has it: ``Actual results
may vary.'')
A TPV morphing algorithm averages the amplitude envelopes and frequency
envelopes of corresponding tracks of the two analyzed les to be morphed. It
matches by looking for tracks in each of the analysis les in which the ratio
of analyzed frequency to that le's fundamental frequency is approximately
equal.

277

Windowed Analysis and Transformation

Morphing clearly diers from simply mixing two sources. In practice, some
morphs are more convincing than others. Two tones with a common pitch tend
to morph more naturally than two tones with dierent pitches, for example.
TPV Pitch-Time Changing
By altering the position or extent of the tracks, one can shift the pitch or alter
the duration of the sound (Portno 1978). For example, to stretch the duration,
the TPV interpolates new points between existing points in the amplitude and
frequency arrays. To shrink the duration by a factor of n, The TPV uses every
nth value in reading the amplitude and frequency arrays. In eect, this shifts the
sampling rate. To shift the pitch of a sound but not change its duration, one
multiplies the frequency values assigned to each of the frequency functions by
the desired factor. For example, to shift a sound up a major second, the TPV
multiples each frequency component by 11.892%.
Disintegration and Coalescence in the Gabor Matrix
The QuickMQ program (Berkely 1994) reads TPV analysis les generated by
the Lemur Pro program. It lets users alter this data through a number of algorithms. The Granny algorithm is especially interesting. Granny pokes holes in
the Gabor matrix of a sound, producing a kind of granular disintegration or
decomposition. The parameters of granular decomposition are as follows:
1. Maximum percentage of zeroing [0, 100]
2. Entropy raterate of decomposition [1, 10]
3. Distribution [random, loud tracks rst, soft tracks rst]
4. Rest in peace (RIP) switch [on, o ]
The maximum percentage of zeroing determines how many zero magnitudes
can be written to a track once it is chosen for decomposition. The percentage
refers to the total number of frames in the analysis le. If the le has three
hundred frames, for example, and the user enters a maximum percentage of
zeroing of 50%, then any track selected for decomposition may be zeroed with
between zero and one hundred and fty frames of zero magnitude. The entropy
rate determines how quickly maximum decomposition settings occur in the le.
Lower values result in more immediate decomposition. The distribution setting
lets the user specify whether louder or softer tracks should be decomposed rst,

278

Chapter 6

or if tracks should be picked at random for decomposition. The RIP switch

directs the algorithm not to decompose tracks that have already been decomposed. In practice, RIP mitigates against the decomposition process.
My experiments show that a disintegration eect can be achieved with the
following ranges:
1. Maximum percentage of zeroing: 3080%
2. Entropy rate: 37
3. Distribution: loud tracks rst
4. RIP switch: o
The eect is imperfect, suering from unwanted transient artifacts. Tracks
suddenly recommence where a zeroed out segment stops. This can be ameliorated by manually editing these transients on a micro time scale. Clearly this
granular decomposition algorithm could benet from a ``windowed zeroing''
that allows tracks to fade in and out over a few milliseconds instead of cutting
in and out abruptly.
This same decomposition algorithm creates its inverse, a coalescence eect,
by simply time-reversing the input le before submitting it to the algorithm, and
then time-reversing the output after processing.
Deterministic Plus Stochastic TPV
The TPV handles static or smoothly changing tones best. TPVs can also analyze and resynthesize many inharmonic sounds, including bird songs and tuned
percussion tones. Since the TPV is based on Fourier analysis, however, it must
translate noisy and inharmonic signals into combinations of periodic sinusoidal
functions. For noisy signals particularly this can be costly from a storage and
computational standpoint; to synthesize a simple noise band requires an everchanging blend of dozens of sine waves. Storing the control functions for these
sines lls up a great deal of memory spacemore than ten times as many bytes
as the original sound samples. Resynthesizing the sines demands a great deal
of computation. Moreover, since the transformations allowed by the TPV are
based on a sinusoidal model, operations on noisy sounds often result in clusters
of sinusoids which have lost their noisy quality.
Researchers have extended the TPV to make it more eective in handling
noisy signals. Xavier Serra (1989) added ltered noise to the inharmonic sinusoidal model in spectral modeling synthesis or SMS (Serra and Smith 1990).

279

Windowed Analysis and Transformation

SMS reduces the analysis data into a deterministic component (prominent

narrowband components of the original sound) and a stochastic component.
The deterministic component tracks the most prominent frequencies in the
spectrum. SMS resynthesizes these tracked frequencies with sine waves. SMS
also analyzes the residue (or residual ), which is the dierence between the deterministic component and the original spectrum. After analyzing the residual
SMS approximates this stochastic component through a collection of simplied
spectrum envelopes. The resynthesis passes noise through a set of lters controlled by these envelopes.
Musicians can modify the stochastic part through graphical operations on
envelopes. However, the perceived link between the deterministic and stochastic
parts is delicate; editing them separately may lead to a loss of recognizable
connection.
Experiences with TPV Transformations
My experience with the TPV has been with the MacOS programs Lemur Pro,
QuickMQ, PAST, and MarcoHack, which share a common le format. Lemur
Pro performs the TPV analysis and resynthesis for all the programs, letting the
user adjust the analysis parameters, including the window and FFT size, hop
size, bandwidth of the analyzing lter, and sidelobe attenuation of the Kaiser
window. Once the analysis is done, Lemur Pro oers operations such as track
editing (select, delete), time-scale modication, spectrum multiplication and
addition (shifting), and track-based amplitude modication.
The PAST program is based on a model of auditory perception. It extracts
additional information from the Lemur analysis data, and generates control
functions for fourteen features: amplitude, spectral envelope, onset spectral envelope, spectral density, onset asynchrony, spectral distribution, amplitude modulation range, amplitude modulation rate, onset amplitude modulation range,
onset amplitude modulation rate, frequency modulation range, frequency modulation rate, onset frequency modulation range, and onset frequency modulation
rate (Langmead 1995). Users can then edit envelopes for these features and
resynthesize the altered sound.
QuickMQ also reads Lemur analysis data and oers a number of modications, including granular decomposition in the time-frequency plane, brightness enhancement, spectral rotation, track smoothing, gain changing, dynamic
rotate above and below a specic frequency, spectrum expansion, falling or
rising spectra, spectral inversion, ``scrunch'' spectrum, transpose spectrum, and

280

Chapter 6

vibrato spectrum. QuickMQ is unfortunately restricted to the older, much slower,

Macintosh Quadra processors. While this discouraged casual experimentation,
I was able to achieve good results through systematic testing.
MarcoHack (Pranger 1999) implements spectral shaking, stretching, and
tracing (Wishart 1994). Spectral shaking adds random deviations to the amplitude and frequency traces of a TPV analysis le. Spectral stretching multiplies
the analyzed frequencies by a constant or variable quantity, and, in this implementation, spectral tracing discards low-amplitude tracks in the analysis.
Assessment of the TPV
The tracking phase vocoder converts the analysis data from the cells of the
Gabor matrix to a set of time-varying tracks. An operation on one track does
not aect other tracks. Manipulations of short tracks, or the segmentation of
long tracks into shorter ones by granular decomposition, result in microsonic
transformations.
Setting the parameters of the analysis is more critical in the TPV than in the
regular PV. This makes planning and experimentation mandatory. Certain
operations, like time-stretching of an instrumental tone, may not require all
of the TPV's exibility and the ordinary PV will work well. The plain TPV is
limited to a sinusoidal model, which is unsatisfactory for noisy short-time
musical signals. For these, however, extensions such as stochastic and transient
analysis-resynthesis add to the TPV's promise.

The Vector Oscillator Transform

The vector oscillator transform, or VOT, is part of the ancient Turbosynth
program (1988, MacOS) by Digidesign. The operation is worth citing, however,
for its intuitive graphical interface. In Turbosynth, the VOT appears in a menu
as ``Convert To Oscillator.'' It applies a series of windowed spectrum analyses
to a sampled sound, each of which covers a short frame of time. The result is a
series of analyzed waveforms read in sequence by a vector oscillator (Roads
1996). A vector oscillator scans not just a single wavetable, but a sequence of
wavetables, crossfading from one to another. The frequency of the playback
oscillator is adjustable.
Figure 6.5 shows a vocal fragment and various forms of its VOT. The user
can click on any of the wavetables to edit them in either the time-domain or the

281

Windowed Analysis and Transformation

Figure 6.5 Vector oscillator transform. (a) First 600 ms of the waveform of the spoken
Italian phrase Nient'altro. We see only ``Nient'alt.'' (b) VOT of this phrase with wavetable extracted at at 20 ms intervals. This turns it from samples into a series of thirty
wavetables read by an oscillator at a stipulated fundamental frequency. (c) The same set
of wavetables stretched in time by a factor of 2.5. In eect, the oscillator takes longer to
crossfade between waveforms. (d) A click on a button converts the waveform view into a
frequency-domain view, with sixty-four harmonics per wavetable.

282

Chapter 6

frequency-domain. The conversion from time-domain to frequency-domain and

back is instantaneous. The temporal distance between the wavetables can be
made longer or shorter. A longer distance means that the sound sustains longer
on a given wavetable and crossfades more slowly. Users can rearrange the order
of the wavetables arbitrarily, leading to scrambling eects.
Assessment of the VOT
The sound quality of the resynthesis produced by the VOT is marked by buzzy
or resonating artifacts. This is not surprising given that it is not a phase vocoder
but more of a ``frequency-domain granulator.'' The VOT's strong point is its
interface with the microtemporal domain. Although developed in the 1980s, its
capacity to edit each frame remains unique to this day. The openness of the
editing (one can insert any waveform or short-time spectrum into a frame) is
exemplary. One could imagine a similar interface for the Gabor transform,
where each cell or groups of cells in the Gabor matrix could be edited. Sonographic synthesis systems approach this ideal.

Wavelet Analysis and Resynthesis

Wavelet analysis measures a signal's frequency content within a window of
time. As opposed to the Gabor matrix and the STFT, with their uniform timefrequency regions (gure 6.6a), wavelet analysis divides the TF plane into
nonuniform regions (gure 6.6a). Low-frequency regions may be long in time
and narrow in frequency range, while high-frequency regions may be short
in time and wide in frequency range. So a cymbal crash remains invisible to
a ``slow'' (low-frequency) wavelet, but will be detected by a burst of ``fast''
wavelets. Prior to wavelet analysis, a family of methods called constant Q
analysis also implemented lter banks in which the bandwidth varied proportionally with frequency (see Roads 1996). In recent years, however, the wavelet
transform has largely supplanted research in traditional constant Q techniques.
Wavelet analysis associates each region of the TF plane with a time-domain
signal (the wavelet) which can be thought of as the impulse response of a
bandpass lter (Vetterli and Herley 1992). Wavelets, scaled in duration and
frequency and translated in time, form an analysis basis for the analyzed signal.
When they are added together in prescribed amounts, certain classes of signals
can be reconstructed.

283

Windowed Analysis and Transformation

(a)

(b)
Figure 6.6 Comparison of Fourier and wavelet time-frequency grids. (a) Uniform
Gabor matrix/STFT grid. (b) Nonuniform wavelet grid.

284

Chapter 6

The wavelet transform (WT) is a multiresolution representation. That is,

it decomposes a signal by successive approximation, starting from a coarse
version and adding time-frequency details on multiple scales. The WT maps
sounds onto a time-frequency grid. The center of each grid is the mean time of
occurrence and the spectral centroid. In music analysis with the WT, one sets
up the grid according to the goals of the analysis and distorts the grid according
to the goals of the resynthesis.
Musical applications often use the Morlet wavelet, which happens to have a
Gaussian envelope. Thus the wavelet is similar to the grain discussed in chapter
3, and to the windowed segments of the STFT discussed earlier in this chapter.
Unlike a grain, however, no matter what frequency a wavelet contains, it
always encapsulates a constant number of cycles. This implies that the size
(duration) of the wavelet window stretches or shrinks according to the frequency being analyzed, referred to in the literature as dilation and contraction.
The implication of a variable window size is that the WT trades frequency resolution for time resolution at high frequencies, and trades time resolution for
frequency resolution at low frequencies.
History of Wavelet Analysis
The term ``wavelet'' and its French equivalent ``ondelette'' were used in early
twentieth century physics to describe the packets of energy emitted by atomic
processes (Crawford 1968). Modern wavelet theory brings together a cluster of
ideas imported from pure mathematics, computer vision, image processing, and
signal processing (Torresani 1989). Yves Meyer (1994) has traced seven dierent origins for modern wavelet theory.
The modern WT was originally developed in 1983 by Jean Morlet, a geophysicist working for Elf Aquitaine, in collaboration with Alex Grossmann of
the Centre Nationale de Recherche Scientique (CNRS) in Marseille (Meyer,
Jaard, and Rioul 1987; Risset 1999). It was applied immediately to problems
in acoustics and image processing (Dutilleux, Grossmann, and KronlandMartinet 1988; Kronland-Martinet and Grossmann 1991; Boyer and Kronland-Martinet 1989; Kronland-Martinet 1988). The wavelet transform was a
slow and computationally intensive operation until 1985, when S. Mallat and I.
Daubechies presented a recursive algorithm for what is called the fast wavelet
transform or FWT (Mallat 1988, 1989, 1998). Meanwhile, engineers and physicists elaborated the mathematics of the WT, and found additional applications

285

Windowed Analysis and Transformation

(Strang 1989; Evangelista 1991; Kussmaul 1991; Vetterli and Herley 1992).
Today wavelet theory is one of the most extensively researched subjects in all of
signal processing.
Operation of Wavelet Analysis
A casual reference to ``measurable functions'' and L 2 R can be enough to make an
aspiring pilgrim weary. (Kaiser 1994)

Wavelet theory is buried in mathematics. Part of the diculty for the nonmathematician is that its foundations are expressed in terms of linear algebra,
mappings between vector spaces, and the theory of generalized functions. A
large portion of the literature is devoted to projecting the extent of wavelet
theory and its abstract relationships to other branches of mathematics. These
abstractions are far removed from applications in musical signal processing.
Only the most basic tenets of wavelet theory can be explained using simple
mathematical concepts.
The Grossmann-Morlet wavelet equation (Meyer, et al. 1987; KronlandMartinet and Grossmann 1991) centered on point b in time, can be dened as:


1
tb
Ca; b t p c
; a > 0; b A R
a
a
The variable a is the scale factor. The wavelet function Ca; b t oscillates at the
p
frequency 1 a. When the scale is very small, the rst factor in the equation,
p
1= a, tends toward 1, while the time interval contracts around b as t b=a
tends toward t b.
The wavelet transform is:



1
tb
Sa; b ca; b tst dt p c
st dt
a
a
where c represents the complex conjugate. In eect, the WT multiplies the input signal st by a grid of analyzing wavelets, bounded by frequency on one
axis and by time scale factor on the other. This multiplication process is
equivalent to convolving with a bandpass lter's impulse response. Dilation of
this impulse response corresponds to an inverse frequency scaling. Thus, the
duration of each wavelet corresponds to the center frequency of a lter. The
longer the wavelet, the lower is its center frequency. The output of the WT is a

286

Chapter 6

two-part spectrum, with one part representing the modulus or magnitude at a

given frequency and the other part representing phase.
The frequency scale of the analysis grid can be logarithmic. This means that
the frequency of each analyzing wavelet is related to the others by a logarithmic
musical interval such as a fth or a third, depending on the way the system is
set up. The use of a logarithmic scale is not mandatory, however, since the WT
can be aligned on an arbitrary frequency scale. In many papers on wavelets, the
analyzing grid is assumed to be dyadic, which can be interpreted as constant Q
ltering with a set of octave-band lters (Vetterli and Herley 1992).
Wavelet Resynthesis
In terms of matrix algebra, the inverse wavelet transform is the transpose of the
analysis. In practical terms, wavelet resynthesis can be carried out in two ways:
overlap-add and oscillator-bank additive (Boyer and Kronland-Martinet 1989).
Each method lends itself to certain types of transformations. In the case of the
overlap-add (the so-called granular) method, one needs as many oscillators as
there are overlapping wavelets. In additive resynthesis, the number of oscillators is constant, since each frequency component is assigned its own oscillator.
Basis Functions
A time-domain signal can be represented as a family of impulses at successive
time points, while a frequency-domain signal can be represented by a family of
sinusoids of varying frequency. In both domains there is an orthogonal family
of functions that represent any arbitrary signal; these are the basis functions.
One of the main points of wavelet theory is that any signal can be decomposed
by means of dierent basis functions. Theory imposes only a few constraints on
this choice. For example, the wavelets must have nite energy and have no DC
bias. Each family of wavelets determines a particular type of analysis on an
audio signal.
In the time domain, an audio signal can be represented by a pair of basis
functions that generates a series of scaled impulses dt
X
at dt
xt
t

In the Fourier frequency domain, the same signal can be represented by a

pair of basis functions that generate a sum of scaled sinusoids:

287

Windowed Analysis and Transformation

xt

at cos2p ft

Wavelet theory represents signals by the weighted sum of a family of basis

functions that are dilations, contractions, and translations of an initial wavelet
function Ct.
X
bjk C2 j t k
xt
j; k

where b is a scaling function, C is a wavelet function, j is the scale factor, and k

is the time translation or delay.
In the terminology of wavelet theory, the support of a wavelet is the region
over which it is nonzero. Wavelets that are nonzero over a narrow range are
said to have a compact support or nite length (Vetterli and Herley 1992). One
can also stipulate a family of wavelets that is nonzero for all k; such a family
has extended support or innite length. If the support of a wavelet family is
compact, it is straightforward to compute the WT.
The Fast Wavelet Transform (FWT)
Direct evaluation of the wavelet transform is computationally intensive. In
1985, S. Mallat and I. Daubechies presented a recursive algorithm for the fast
wavelet transform (FWT). It can calculate the forward and inverse wavelet
transform of a signal in a time period that is of the order n  log2 n, where n is
the number of samples analyzed. At each level of recursion, the algorithm
extracts the details at that level of resolution, and constructs a coarser (less
detailed) version of the signal for analysis at the next level. This process repeats
on successively coarser representations of the signal, until only the average value
of the signal remains. The inverse transform works in reverse. It adds detail to
the signal's coarse representation from higher and higher levels of resolution.
The FWT is a subset of a more general transform, the wavelet packet transform (WPT), developed by R. Coifman. In the FWT, the sampled data passes
through lowpass and highpass lters with complementary bandwidths. The
highpass output is the detail, and the lowpass output is the approximation. The
output of the approximation lter, decimated by a factor of two, can be fed to
another pair of wavelet lters, identical to the previous pair, to yield yet
another set of detail and approximation coecients. This process of multiscale
analysis can continue until only a small unit data set is left.

288

Chapter 6

Wavelet Display
A byproduct of research in wavelet analysis is the evocative display method
developed by scientists aliated with the Centre Nationale de Recherche Scientique (CNRS) in Marseilles. This visualization tool can be thought of as a
traditional spectrum plot projected in time and ipped on its side. In the modulus display, time projects horizontally, and frequency projects vertically with
the low frequencies on the bottom and the high frequencies on the top. The
dierence between the sonogram plot and this wavelet plot is their pattern of
time localization. Short wavelets detect brief transients, which are localized in
time, sitting at the apex of a triangle. Kronland-Martinet et al. (1987) shows the
wavelet plot of a delta function; its wavelet display clearly projects a triangle on
the frequency-versus-time plane, pointing to the locale of the impulse. Long
wavelets detect low frequencies; they sit at the base of the triangle, spread out
(blurred) over time.
The triangle of the delta function is the wavelet's domain of inuence in time.
The domain of inuence for frequencies is a constant horizontal band, as in the
spectrogram. The darker the band, the stronger the magnitude within that frequency range. Using a log scale for the dilation axis allows a greater range of
scales to be observed, important in audio applications where frequencies can
vary over several orders of magnitude.
A voice is a set of transform coecients with xed dilation parameters. Thus
a voice in some ways corresponds to a frequency band in an equalizer. If the
frequency grid is aligned to a musical interval, the modulus projects a strong
dark indicator when the input signal contains that interval.
A plot of the phase spectrum is sometimes referred to as the scalagram.
The scalagram yields interesting details about the phase transitions within a
given signal, such as the onset of transients and the nature of the modulations
within the signal (Kronland-Martinet, et al. 1987; Kronland-Martinet, et al.
1997).
Transformation of Sounds Using Wavelets
Once a sound signal has been analyzed with the WT, one can alter the analysis
data in order to transform the original signal. This section describes the various
transformations.

289

Windowed Analysis and Transformation

Substituting a New Synthesis Wavelet

One possibility oered by the WT is that of substituting a dierent wavelet
function for the resynthesis than was used for the analysis. Using Mallat's
smooth wavelet for the analysis, for example, and substituting Haar's boxcar
wavelet for the synthesis introduces a noisy distortion eect. From a technical
standpoint, this is a straightforward operation.
Wavelet Filtering
Another direct transformation is to lter by suppressing certain frequency
channels in the resynthesis. Since wavelet representations involve both translation and scale parameters, one can impose amplitude envelopes at dierent
scales, so that the frequency content of the signal changes with time. The logarithmic spacing of the voices makes it possible to extract chords from the sound
by setting all but a few voices in each octave to zero (Boyer and KronlandMartinet 1989). When this technique is applied to the speaking voice, for
example, it gives the impression of a person talking ``harmonically.'' Crosssynthesis possibilities are also suggested by this transformation. KronlandMartinet notes that cross-synthesis can be achieved by resynthesizing from the
modulus of one sound and the phase of another.
Pitch-Time Changing with the Wavelet Transform
Pitch-time changing with the WT can be accomplished using complex-valued
signals and wavelets, since it involves stretching the grid of scale and translation
values to achieve the desired duration, and then multiplying the phase of the
values so that the resulting pitch is unchanged. To shift pitch by a constant
factor, one multiplies the phase values of the analyzed wavelets by this constant. To dilate or contract the timebase while keeping the pitch the same, one
can expand or compress the point of overlap of the wavelets in resynthesis.
Comb Wavelet Separation of Noise from Harmonic Spectrum
The comb wavelet transform, developed at the University of Naples Federico
II, sorts transients, unpitched sounds, and pitch changes from quasiperiodic
signals (Evangelista 1992, 1997; Piccialli et al. 1992). The comb WT starts from
a windowed segment of sound. The algorithm estimates the fundamental pitch

290

Chapter 6

period, and ts a comb lter to the segment with peaks aligned on the harmonics of the fundamental. The comb lter sifts out the energy in the harmonic
spectrum. The algorithm then performs a wavelet analysis on this ``clean'' harmonic signal. When the inverse WT is subtracted from the original signal, the
residual or ``dirty'' part of the signal remains. The dirty part includes the attack
transient and the details that give the sound its identity and character. Once the
clean and dirty part are separated, one can perform a kind of cross-synthesis by
grafting the dirty part of one sound into the clean part of another. This type of
separation is similar in conceptthough not in implementationto the technique used in the spectral modeling synthesis of Serra (1989).
Evangelista and Cavaliere (1998) extended the comb wavelet technique to
handle the case of inharmonic partials, warping the frequencies of the analysis
to adapt the pitch-synchronous comb wavelet lters to unequally spaced partials. With this technique they were able to resolve separately the hammer noise
and the resonant components of a piano tone. Frequency warping also allows
for transformations such as detuning the microintervals within a complex
sound.
Other Wavelet Transformations
Other transformations include altering the geometry of the frequency grid, such
as multiplying by or adding a scaling factor to all the frequencies in resynthesis.
Cheng (1996, 1997) described an application of wavelets to make a spectral
``exciter'' eect in which high frequencies are boosted and an additional octave
of high frequency information is extrapolated from existing lower frequencies.
According to the author, this algorithm worked best on broadband transient
percussion sounds, but was not well adapted to speech, where the added frequencies were perceived as articial.
Experiences with Wavelets
Wavelet-based software for audio applications is rare at present. For the
MacOS platform there is WaveLab, a library for the Matlab environment
developed at the Department of Statistics at Stanford University and the
National Aeronautics and Space Administration. WaveLab was designed for
wavelet analysis, wavelet-packet analysis, cosine-packet analysis, and matching
pursuit. It also provides a library of data les, including articial signals as well
as images and a few brief sampled sounds. As is, however, the WaveLab pack-

291

Windowed Analysis and Transformation

age is not well suited for musical experimentation. A great deal of additional
programming would be required for serious experimentation in musical sound
transformation.
Soniqworx Artist by Prosoniq is an audio editor with signal processing eects
for MacOS. The eects, written by Stephan Sprenger, include wavelet-based
transformations. One of these, wavelet signal reduction, discards all but a
stipulated percentage of the analysis wavelets, leaving only the strongest components. It is conceptually analogous to the spectral tracing eect described in
the section on the phase vocoder. Instead of sinusoidal basis functions, however, it uses wavelet basis functions. Owing to the change in basis function from
a sinusoid to the jagged Daubechies second wavelet, the sonic eect is quite
dierent from spectral tracing (gure 6.7).
Synthetic Wavelets
One could synthesize a stream of wavelets without an analysis of an existing
sound. Malvar wavelets (Meyer 1994), for example, bear a strong resemblance
to the FOF particles and to the grains with a quasi-Gaussian envelope. My
grainlet technique (described in chapter 4), produces grains whose duration is a
function of frequency, like the usual wavelet families. Wickerhauser (1994)
proposed that a single wavelet packet generator could replace a large number of
oscillators. Through experimentation, a musician could determine combinations of wavelet packets that produce especially interesting sounds. It could also
be possible to reproduce the sounds of traditional instruments by decomposing
an instrumental sound into wavelet packet coecients. Reproducing the note
would then require reloading those coecients into a wavelet packet generator
and playing back the result. Transient characteristics such as attack and decay
could be controlled separately (for example, with envelope generators), or by
using longer wave packets and encoding those properties into each note.
Assessment of the Wavelet Transform
Almost every author seems to have a favorite basis function, depending on their area of
interest and background. (Navarro et al. 1995)

Traditional Fourier methods measure the average energy across a window

whose duration remains constant, no matter what frequency component is
being analyzed. This tends to delocalize the view of the onset time of high-

292

Chapter 6

Figure 6.7 Wavelet signal reduction compared with spectral tracing. (a) Original
waveform of spoken voice saying ``pulse code modulation.'' (b) Resynthesized waveform
after 90% of the weaker wavelets have been discarded. (c) The uctuation indicated by
the ` symbol in (b) blown up. It has the shape of the second Daubechies wavelet, which
was the basis function of the analysis. (d) The same region of the sound le covered by
(c) but resynthesized after spectral tracing, which discarded all but the strongest 1% of
sinusoidal components.

293

Windowed Analysis and Transformation

frequency transients. In contrast, the WT oers a multiresolution view of a

musical signal, since ne temporal analysis is resolved with short, high-frequency
wavelets, while long, low-frequency wavelets resolve ne frequency dierences.
Thus the WT is well suited for the study of transients or onset times in musical
signals, because it exhibits remarkable time sensitivity at high frequencies
(Dutilleux, Grossmann, and Kronland-Martinet 1988). In the past, FFT-based
methods had a computational advantage over wavelet methods. By now,
engineers have developed optimizations for many types of wavelet transforms
(Shensa 1992).
Torresani (1995) pointed out how group theory operations (translation,
expansion, compression, rotation) can be applied to signals in the wavelet
domain. As Arb noted in 1991, however, an arbitrary geometrical transformation of a time-frequency analysis is not guaranteed to produce an interesting
sonic result. Meijer has argued:
Unless the mathematical properties of an exact lossless and/or orthogonal mapping and its
inverse reconstruction are considered essential, there is no convincing reason to use wavelets in sound synthesis and analysis. For instance, in auditory perception almost nothing is
exact, even though the localization properties in time and frequency are very important.
There is also no reason why the resonance properties of the basilar membrane would be
best described using wavelets, since the constant-Q property approximates only part of the
membrane. Constant bandwith, as suited to xed-width time windowing, better approximates the admittedly often less important low-frequency part. . . . Furthermore the human
hearing system is known to be nonlinear in a number of ways, so even without knowing how
to exploit that one can argue that a linear mapping is not going to be an optimal mapping
for human auditory scene analysis. (Meijer 1998)

At present, the vast majority of wavelets in audio are not concerned with
artistic transformations of sound. The principle applications are utilitarian:
1 Audio data reductionfor transmission in broadcast and network media.
Wannamaker and Vrscay (1997) report compression ratios of 3 : 1 with
``generally satisfactory reconstruction.'' The results do not preserve high
delity. Barnwell and Richardson (1995) criticize wavelet techniques for
audio compression, particularly the dyadic wavelet grids based on octaveband lter banks. (See also Erne 1998.)
1 Denoisingremoval of broadband noise from an audio recording, with
``faint birdy noise'' remaining as an epiphenomenon of the denoising process
(Ramarapu and Maher 1997).

294

Chapter 6

1 Detection and classication of transient signals in background noisefor

example, detection and classication of underwater acoustic signals generated by whale clicks, or scratches in vinyl recordings (Grossmann et al.
1987; Learned and Willsky 1993).
1 Display of spectrum featuresusing the properties of the wavelet technique
to isolate specic features of audio signals and project them visually
(Kronland-Martinet, et al. 1987; Newland 1994).
1 Detection and display of internal modulations (Arb and Delprat 1993,
1998; Delprat et al. 1990; Kronland-Martinet, et al. 1997).
1 Analysis of turbulence and chaosthe wavelet transform can serve as an
alternative to the Fourier decompositions in analyzing turbulence (Liandrat
and Moret-Bailly 1990).
1 Analysis of performed rhythmsTait (1995) applied wavelets to the display
of rhythmic patterns in jazz.
1 Sound texture matchingBar-Joseph et al. (1999) analyzed stationary or
statistically constant sound textures (rain, waterfalls, crowds of people,
machine noises, etc.) using a multiresolution wavelet approach. They could
then synthesize similar textures using granular synthesis.
The mathematical conditions that must be satised to create a wavelet function are ``weak'' or relatively unconstrained. This oers great freedom in the
choice of wavelet. Specic wavelet functions can be devised to detect special
features in an audio signal. Kronland-Martinet, et al. (1997) imagined the
decomposition of a signal in order to detect a specic word pronounced with
dierent pitches and durations. This exibility is seen in some of the analysis
applications just mentioned. The musical potential of feature-based searching
could be great, especially when coupled with selection and transformation.
Wavelets have been proposed as a component in the modeling of weak turbulence or deterministic chaos (Arneodo et al. 1995). In certain forms of chaos,
the pattern of turbulence is similar (but not identical) on dierent time scales.
Since wavelets oer a multiscale representation, they can be combined with
multifractals. These are mathematical models of intermittence in the small scale
of turbulence. This combination could create multilevel models of deterministic
chaos. Fractal systems can model the degree of regularity at each scale, leading to the concept of a spectrum of singularities (irregularities). This is a rich
territory of scientic exploration which may well have artistic potential. (See
Robindore 1996a.)

295

Windowed Analysis and Transformation

The wavelet paradigm has generated great interest since its introduction in
the 1980s. In the larger context of the past several decades of signal processing,
however, wavelets and multiresolution lter banks have oered a limited range
of musical applications. Fourier and Gabor techniques remain dominant, and
powerful applications such as fast convolution have stayed within the Fourier
camp. Masri, et al. (1997a) suggest that the logarithmic frequency scale of the
wavelet transform is an advantage in detecting pitch, but that the linear
frequency scale of the Fourier and Gabor transform is an advantage in analyzing linear-space harmonics. Nevertheless, it seems clear that the full scope of
wavelet techniques has not yet been exploited in the musical domain.

Gabor Analysis
Expansion into elementary signals is a process in which Fourier analysis and time description are two special cases. (Gabor 1946)

As chapter 2 recounts, in the 1940s, Dennis Gabor proposed a comprehensive

approach to the analysis of sound, authoring a series of important scientic
papers combining insights from quantum physics and psychoacoustics with
practical experiments (Gabor 1946, 1947, 1952). In these papers, he suggested
that the representation of a signal purely in terms of its time series (e.g., the
sequence of samples of a discrete signal) or purely in terms of its frequency
content (Fourier expansion) were both extreme cases of a wide range of signal
expansions. They show high locality (precision) in one domain, but no locality
at all in the other. For the analysis and manipulation of sound, good locality in
both time and frequency is desirable. Gabor showed that any sound could decompose into a combination of elementary acoustical quanta. A plot of the
analysis data on a time-versus-frequency rectangular grid is the Gabor matrix
or lattice. Each quantum has a nite extent in time and in frequency. Gabor
proposed a smooth Gaussian window for the quantum, a natural choice since
the spectrum of a Gaussian window resembles a similar bell-shaped curve. In
this case, a locality in the time domain corresponds to the locality in the frequency domain.
Theory of the Gabor Transform
Gabor's design for sound analysis was similar to the short-time Fourier transform (STFT) which applies a sequence of localized Fourier transforms on

296

Chapter 6

xed-duration windowed segments of a source signal. The STFT emerged out

of research in speech analysis and resynthesis in the 1960s and 1970s at the Bell
Telephone Laboratories. During this period, Gabor's pioneering work was
neglected.
Gabor's papers described an iterative approximation method to calculate the
matrix. By 1966, Helstrom had shown how to recast Gabor's analysis and
resynthesis approximation into an exact identity by turning the elementary signals into orthogonal functions. Bacry, Grossman, and Zak (1975) and Bastiaans (1980, 1985) veried these results, and developed analytic methods for
evaluating the matrix.
The analysis procedure that results in this time-frequency representation is
the Gabor transform (GT). At the core of the GT is the gaboret (Arb and
Delprat 1993), a complex exponential signal limited in duration by a Gaussian
window. The GT multiplies the gaboret at dierent frequencies by a succession
of windowed segments of the input signal. These multiplications generate
the Gabor matrix, also known as the modulus of the GT (gure 6.8). Since
the gaborets are complex, the GT also produces a representation of phase,
called the phasogram. Researchers have used this representation to reveal
phase periodicitiesindicators of modulation laws hidden in the input signal
(Kronland-Martinet et al. 1997). See Arb 1991 for a discussion of the dierences between the phase vocoder and the GT.
A drawback of Gabor's original design was the lack of pure orthogonality.
This is because of the smoothness of the Gaussian window, which does not trail
perfectly to zero at its extrema, leaving a small DC component. When using the
same window in the resynthesis, the result is unstable, leading to spikes in the
output. Researchers have spent great eort on the construction of biorthogonal
or dual functions, a set of compensating windows that correct for the lack of
orthogonality of the Gaussian window (Helstrom 1966; Bacry, Grossman, and
Zak 1975; Bastiaans 1980, 1985; Navarro et al. 1995a). Bastiaans and Geilen
(1996) showed how the discrete Fourier transform, in conjunction with the
discrete Zak transform, can calculate the discrete Gabor transform in a sumof-products form, which leads to fast algorithms.
The owering of wavelet theory since the 1980s has had a strong inuence on
the modern theory of the Gabor transform, leading to a new characterization of
the gaborets, also called the Gabor wavelets or gaborettes (Torresani 1995). This
has led to a reformulation called the multiresolution or discrete Gabor transform
(DGT). In the DGT, the complex spectrogram of a signal is dened as the
Fourier transform of the product of the signal and the shifted and complex

297

Windowed Analysis and Transformation

Figure 6.8 Schema of the Gabor transform. The GT multiplies a segment of a source
signal by a complex Gaboret, translated in frequency and time according to the resolution of the Gabor matrix. The energy occupies cells in the modulus of the GT, and the
phase information appears in the phasogram. To resynthesize the source, each cell of the
matrix is convolved with the reproducing kernel, the dual of the complex Gaboret.

conjugated version of a window function. The complex spectrogram is determined by its values on the points of the Gabor lattice. In recent years, mathematicians have pushed the theory of the GT beyond the Gaussian gaborets
to account for larger classes of particles: general subgroups on the TF plane
(Torresani 1995).
Properties of the Gaussian Window Used in the GT
Harris (1978) compared the properties of twenty-three dierent window shapes,
including the Gaussian window used in the GT. According to Harris, two
properties of a window are paramount in spectrum analysis. One is its highest
side lobe level or HSSL. To maximize frequency detectibility, this should be low
relative to the central lobe. The other is its worst case processing loss. The
WCPL indicates the reduction in output signal-to-noise ratio as a result of
windowing and worst case frequency location. It should, therefore, be low, or
the quality of the analysis will be compromised.

298

Chapter 6

The HSSL of the rectangular Dirichelet window is only 13 dB, a poor gure that results in noise-cluttered spectral analysis. The WCPL of the Dirichelet
window is 3.92 dB. In comparison, the Gaussian window is considerably more
selective, with a highest side lobe level of 42 to 69 dB, depending on the
peakedness of the envelope. Its WCPL ranges from 3.14 to 3.73 dB.
Other windows outperform the Gaussian window in these properties. A narrow Kaiser-Bessel window has an HSSL of 82 dB, while a wide Hanning
window has a WCPL of 3.01 dB. These cases are somewhat opposite to each
other, however, and the Gaussian window remains a good compromise between
the two.
Musical Applications of the Gabor Transform
Daniel Arb, working at the Laboratoire de Mecanique et d'Acoustique (Centre Nationale de Recherche Scientique, Marseilles), was the rst to apply a
digital Gabor transform to the analysis and resynthesis of musical sounds
(Arb 1990, 1991; Risset 1992). He was soon joined by his colleague Nathalie
Delprat. One of their rst results was a robust time compression and expansion
operation in which pitch remains constant. They also applied the GT to other
musical applications including frequency transposition, phase manipulations,
cross-synthesis of speech with other sounds, and modifying the vibrato of a
sung vocal tone while keeping other characteristics intact. They also separated
the noisy, inharmonic parts of a sound from the harmonic parts (Arb 1990,
1991; Arb and Delprat 1992, 1993, 1998).
Kronland-Martinet et al. (1997) combined the Gabor transform with the
innovative display techniques developed for the wavelet transform in Marseilles. Their plots show the phase structure of signals, which is normally hidden
from view. According to the authors, this phase information can be used to
make frequency estimations more accurate. They also applied the GT to the
estimation of the amplitude and frequency modulation laws of musical instrument tones. (See also Delprat et al. 1990.) This work may lead to a more complete analysis of musical tones, one that quanties the energy at each point in
the time-frequency plane, but that also accounts for its internal modulations.
This separation opens up the possibility of modifying the energy distribution
independent of the modulation, or vice versa, which is impossible with standard
Fourier techniques.
Leigh Smith (1996) proposed a Gabor wavelet representation of performed
rhythms (pulsations between 0.1 to 100 Hz in frequency) as a model for the

299

Windowed Analysis and Transformation

human perception of musical phrase groupings. The analyzer indicated the

locations of accents, and could infer musical phrase structure from these cues.
Assessment of the Gabor Transform
Advances in the theory of the Gabor transform proceed in two directions. The
rst direction consists of elaborations on the plane of pure mathematics: newer
theorems, higher abstractions, and more far-ung links to other branches of
mathematics. Here the signal is quite secondary in importance, indeed, there
may be no signal. The second direction extends the range of applications of the
GT in the analysis and transformation of sound. For musicians, the rst direction is of interest only insofar as it may foster the second.
Like the short-time Fourier transform and the wavelet transform, the GT is
limited in resolution by a fundamental uncertainty principle. The sponge of the
reproducing kernel blurs any manipulation of the Gabor matrix.
As Jean-Claude Risset (1992) has observed, the GT and the wavelet transform have many points in common. They are both concerned with elementary
functions limited in time and in frequency. They each oer modication of the
analysis data using an arbitrary function applied to the reproducing kernel.
They both allow the possibility of resynthesis from a discrete analysis grid. In
musical applications, Risset has noted:
[The Gabor transform] can produce excellent results in the processing of hyperbaric
speech and in slowing down a spoken recording by a factor of more than 100 whilst preserving intelligibility (on the condition of only slowing down the vocalized parts of the
signal). (Risset 1999b)

It is clear that the GT is a potent means of sound transformation. The work

of the Marseilles group has been especially important in adapting wavelet concepts to the GT. Fortunately for us, their focus is on acoustical problems.
Hopefully this research will eventually be made into widely available software
tools.

Summary
Where the telescope ends, the microscope begins. Which has the grander view?
Hugo 1862)

(Victor

300

Chapter 6
By allowing noise to make inroads into musical sound, Varese accelerated a trend toward
the shattering of traditional musical language. Together with the inadequacy of neoserialism, this resulted in a fragmentation of musical language and a consequent proliferation of composition techniques, very far from a possible theoretical unication. On the
other hand, the increase in methods of acoustic analysis created a situation analogous to
the physics of the microcosmosan imagined unity of sonic phenomena. Here began an
opposition to the continuous wave theory of sound, a granular atomism that was capable of
representing any chaotic state of sound material and was also eective on the plane of
synthesis. (Orcalli 1993)

Beginning in the 1950s, a handful of visionary scientists, acoustical engineers,

and musicians such as Gabor, Meyer-Eppler, Xenakis, and Stockhausen proposed new theories of sound organization based on microsonic particles. Critics
attacked them on the basis that sound quanta were divisible and therefore
not fundamental, thereby missing the point. While there are no fundamental
sound particles, any sound can be decomposed into an innumerable number
of dierent particles, depending on the chosen basis functions. Each such
decomposition makes possible myriad sound transformations. The catalog of
transformations based on window spectrum analysis undoubtedly will continue
to proliferate.

Microsound in Composition

Experiments in Digital Granular Synthesis

Klang-1: Study in Digital Granular Synthesis
Prototype: Study in Automated Granular Synthesis
Compositional Logic of Prototype
Microsound in Compositions by the Author
Polychromatic Timbre in nscor and Field
Clang-Tint
Half-life
Tenth vortex and Eleventh vortex
Microsound Techniques in Works by Various Composers
Barry Truax
Horacio Vaggione
Microsound as a Foreground/Background Texture
Musical Examples by Other Composers
Summary
Scientic tests help us estimate the potential of a technique of synthesis or
sound transformation. They may even suggest how to compose with it, but the
ultimate test is artistic. The aesthetic proof of any signal processing technique is
its use in a successful composition. The rst part of this chapter presents early
experiments in digital granular synthesis, the middle section examples of microsound in my own compositions, and the nal section examples of microsound
in compositions by other composers.

302

Chapter 7

Experiments in Digital Granular Synthesis

Gabor carried out the rst experiments in sound granulation using optoelectronic parts from lm projectors (see chapter 2). Several others, including
Schaeer's colleague Jacques Poullin, built conceptually similar devices based
on magnetic tape technology. In 1959, Iannis Xenakis completed a short granular synthesis study, Analogique B, realized by splicing together hundreds of
analog tape fragments. (See the description in chapter 2.) These heroic, laborintensive experiments gave a glimpse of what might eventually be achieved with
digital technology.
Klang-1: Study in Digital Granular Synthesis
My rst study in digital granular synthesis was Klang-1, realized in December
of 1974. Chapter 3 described the technical environment: a large mainframe
computer running the Music V synthesis language. Klang-1 was a technical test
of three parameters of granular synthesis: (1) the grain envelope, (2) the grain
duration, and (3) the density of grains in time. To prepare, I graphed on paper
two time-varying curves, which indicated the frequency and density of the
grains. From these, I designed a detailed score that specied the starting times
and frequencies of seven hundred and sixty-six grains over a 30-second period. I
then typed this numerical data onto punched cards, one grain per card. The
Music V instrument was a simple sine generator with a Gaussian test envelope.
Music V interpreted each grain as a NOTE statement which fed numbers of the
synthesis instrument.
The duration of every grain was xed at 40 ms, a gure derived from Xenakis's theory (1992, p. 54). The density of the grains ranged from zero to a
maximum of twenty-ve grains per second. The converter hardware xed the
sampling rate at 20 kHz. Grain frequencies in the Klang-1 etude varied from
16.11 Hz to 9937.84 Hz. Between this range, I calculated a 1000-tone scale,
which divided each whole tone into twenty-four subdivisions (Roads 1975), and
approximated the just-noticeable-dierence (JND) thresholds for pitch (Moles
1966).
I did not intend Klang-1 to be anything more than a technical experiment.
Yet I recall vividly the magical impression that the granular sound had on me
as it poured forth from the computer for the rst time.

303

Microsound in Composition

Prototype: Study in Automated Granular Synthesis

The Prototype study incorporated renements derived from the Klang-1 experiment. More than a technical test, it was also a preliminary attempt to compose
with granular sound. The technical renements included:
1. Automated grain generation from high-level cloud specications
2. Use of a quasi-Gaussian grain envelope, with a smooth attack and decay
and an extended sustain, also known as a cosine taper or Tukey window
(Harris 1978)
3. Shorter grain durations (20 ms)
4. Increased grain density
Prototype is an eight-minute granular study, written in April of 1975. The
composition followed a set of systematic procedures that I worked out by hand.
Later, I incorporated these procedures into a larger program for algorithmic
composition (Roads 1976, 1987).
The rst task that I confronted was automating the grain generator. Integral
to Xenakis's conception of granular synthesis was the notion of global stochastic control.
For a macroscopic phenomenon, it is the massed total result that counts . . . Microsounds
and elementary grains have no importance [in themselves] on the scale that we have
chosen. Only groups of grains and the characteristics of these groups have any meaning. (Xenakis 1992)

To realize Prototype, I wrote PLFKLANG, a program in Algol to generate

thousands of grain specications from seven high-level cloud parameters.
PLFKLANG implemented asynchronous granular synthesis (as described in
chapter 3) with a grain waveform that could vary between a sine wave and
bandlimited pulse.
Compositional Logic of Prototype
Prototype is the result of systematic compositional procedures. Involving interaction among fourteen synthesis parameters. These parameters described the
global properties of a sound cloud lled with grains: begin time, duration,
center frequency, center frequency slope (increasing or decreasing), spectral
bandwidth, spectral bandwidth slope (increasing or decreasing), rate of spatial
change, amount of reverberation, and grain density of a set of granular clouds.

304

Chapter 7

Any number of synthesis parameters could interact, meaning that as a

``master'' parameter varied, another parameter linked to it would mirror these
variations. The connection between the parameters was not a simple linear
scaling function, however, but rather a linkage between their degree of order or
disorder. This meant that if a master parameter remained stable (ordered) over
the course of calculating several clouds, then all parameters linked to it would
also remain stable, and vice versa. Another possibility was an inverse link between parameters. In this case, a stable master parameter would cause parameters inversely linked to it change rapidly (disordered).
This type of compositional logic encourages entropic contrasts. By keeping
some synthesis parameters stable, one imposes continuity between successive
sound clouds, while letting other parameters vary more freely.
Behind Prototype are twenty dierent congurations for parameter linkages,
each corresponding to a specic section of the piece. I drew these as a set of
twenty interconnection digraphs. (A digraph, or directed graph, consists of a set
of nodes connected by lines with arrowheads on each line.) Figure 7.1 shows the
rst four of these digraphs. Each digraph corresponded to a specic section of
the piece.
This procedure resulted in an ``Entropy Chart'' that indicated the degree of
order or disorder associated with each parameter in each section. I interpreted
the tendencies indicated by the chart, and developed specications for one
hundred and twenty-three clouds. I typed these numerical specications on
punched cards. The PLFKLANG program read them. A typical specication
was:
SV1

BEG

DUR MU

BETA DEN

DELTA MUSL BESL DESL

1.0

90

99

100

87

30

where BEG is begin time of a cloud, DUR is cloud duration, MU is the initial
center frequency of a cloud, BETA is the initial bandwidth of cloud, DEN is
the grain density, DELTA is the initial amplitude of a cloud, and MUSL,
BESL, DESL represent the time-varying slopes of the MU, BETA, and
DELTA parameters over the course of the cloud, respectively. I then prepared
a graphic score, which plotted each cloud over time. The graphic score also
shows variations in reverberation and spatial panning, which I added later.
I divided the sample calculations into eight one-minute sections. The entire
synthesis process involved sixty-three steps of compilation, calculation, data
transfer, and digital-to-analog conversion over a period of weeks. After assem-

305

Microsound in Composition

Figure 7.1 The rst four digraphs of Prototype. The arrows indicate parameter linkages
(see the text).

bling the eight sound fragments on analog tape, I took this tape to the Village
Recordera sound studio in West Los Angelesfor the nal mixdown on a
Quad-Eight mixing console. There I added electronic plate reverberation (EMT
reverberator with a 2.5 second reverberation time) and spatial panning according to the graphic score.
I played this etude to a number of friends and colleagues as a demonstration
of granular synthesis, and once in a concert in May 1975 at UCSD. Around the
same time, David Cloud, host of the new music program Zymurgy, broadcast
Prototype on the radio station KPFK-FM in Los Angeles. I can hardly imagine
the eect this strange experiment must have had on the unsuspecting listener.

306

Chapter 7

Microsound in Compositions by the Author

Four of the compositions that I made between 1980 and 1999 used microsound
techniques. This section describes these works. Since 1995 I have also realized a
large number of experiments and particle studies. At the time of this writing
(2001), I am assembling this material into new compositions.
Polychromatic Timbre in nscor and Field
Granular synthesis appears briey in two compositions I made in the early
1980s: nscor (1980) and Field (1981). An article on the realization of nscor
appears in (Roads 1985a, 1985b), so this description will be concise. nscor was
realized in the Experimental Music Studio at the Massachusetts Institute of
Technology (MIT). The starting point was a collection of tapes of source material that I had produced over the previous ve years at dierent studios. I
wanted to create a new work in which quite dierent sounds could all function.
This was a deliberate attempt to go beyond the monochromatic approach, or
the use of a single synthesis technique. The goal was to try to expand the range
of sound color toward a polychromatic palette.
To convert, store, and mix dozens of sound lesas we do now so easily
was not possible at the MIT studio during that era. (Chapter 3 describes the
computing environment.) The only practical way to realize the piece was to
assemble it with analog technology. This meant using the studio's mixing console, and the four-track and two-track tape recorders. I mixed and spliced nscor
in the late night hours between March and August of 1980. It premiered at the
1980 International Computer Music Conference in New York, and appeared
on a compact disc released by Wergo in 1987 (Wergo 201050).
Granular synthesis appears in three parts of nscor. In the rst appearance at
0:39 sec, a half-second burst of high-frequency grains, sounding like the asynchronous ringing of a small metal bell, concludes a rapid phrase produced by a
Vosim generator at the Institute of Sonology, Utrecht. Immediately thereafter,
at 0:41 to 0:44, a scintillating cloud of grains in the same register ends the
introductory section. The second appearance is a granular explosion midway
through the work, which is the primary crescendo of the composition. This explosion begins gradually as a sparse scattering of grains at 4:29. It builds with
increasing density and broadening of bandwidth to a sustained peak at 4:36
that abruptly ends at 4:41. Finally, between 5:06 and 5:40, a section constructed
entirely out of thirteen granular clouds serves as a transition between two additive synthesis textures (gure 7.2).

307

Microsound in Composition

Figure 7.2 Spectrum display of granular clouds in nscor by the author, from 5:06 to
5:40. The frequency scale is linear.

Composed in the winter of 198081, Field is another polychromatic composition, in which granular synthesis serves as one instrument in a larger orchestra. Field, like nscor, formed from the ground up, starting with the assembly of
sound objects, building larger structures as the composition continued. Again, I
realized the montage with analog equipment. A major dierence between nscor
and Field was the number of tracks that I had to work with. I now had the use
of a 24-track recorder at a studio in Boston (thanks to a grant from the MIT
Council for the Arts), making it much easier to create an eective montage.
The crossfading ow of musical shapes in Field reects this. The composition
premiered on New Year's Eve 1981, as part of Boston's First Night Festival.
In 1985, Field appeared on a compact disc sponsored by Sony and produced
by the MIT Media Laboratory as part of its inaugural. This CD was reissued
in 1999.
Granular synthesis appears only once in Field, in the form of a granular explosion at 3:383:43. This explosion is the climax of the piece, dividing its two
major sections.
Clang-Tint
Clang-Tint (1994) is not built exclusively from microsounds, but it is a work in
which microsounds are integrated with other sounds. Its point of origin traces

308

Chapter 7

back to a sunny late afternoon, 9 December 1990, following a visit to an exhibition of photographic works by the Starn brothers at the Akron Museum of
Art (Ohio, USA). These works combine prints and large transparencies with
wood, tape, metal pipes, and other media to create three-dimensional sculptures. I was intrigued by the mixture of ``sampled'' imagery in conjunction with
idiosyncratic methods of construction. Inspired by the exhibition, I imagined a
structure for a new composition. Shortly thereafter I received a commission
from the Kunitachi Ongaku Daigaku (Kunitachi College of Music, Tokyo) and
the Bunka-cho (Japanese Ministry for Cultural Aairs). The title Clang-Tint
refers to the notion of ``sound-color.'' The composition starts from sampled
sounds, cut and framed in myriad ways, spatialized throughout, and mixed with
unusual synthetic sounds.
Parts of the work were composed during residencies in Tokyo (1991 and
1994), as well as at my studio in Paris and at Les Ateliers UPIC. At the
Kunitachi school, I was fortunate to have access to the Gakkigaku Shirokan
(musical instrument museum). There I recorded forty-ve instruments, some
two millennia old, some as modern as the Ondes Martenota vacuum-tube
electronic instrument. These recording sessions resulted in a database of ten
hours of sound.
I conceived the composition in four sections: Purity, Filth, Organic, and
Robotic. Each section takes its identity from the sound materials and organizational principles used in it. For example, Purity explores a simple musical
world of sinusoidal waves and harmonies derived from a microtonal scale
created for this composition.
The second movement, entitled Organic, focuses on expressive phrasing. It
combines bursts of insect, animal, and bird calls with pulsar synthesis. Pulsars
appear throughout the composition in dierent forms: pulsating blips, elongated formant tones, and clouds of asynchronous pulsars. For the latter, I rst
generated multiple infrasonic pulsar trains, each one beating at a dierent frequency in the range of 6 to 18 Hz. I then mixed these together to obtain the
asynchronous pulsar cloud.
``Dirty'' sounds make up the sound material of Filth: a morass of crude
waveforms, raw transients, irregular globs and grains, industrial noises, and
distorted tones. Most of the synthetic particles appear in its rst section. After
an intense climax, the nal section consists of soft sounds in layers. Running
throughout most of the section is a recording of burning embers of wood. I
ltered and ring-modulated these natural particles to narrow their bandwidth
to a specic range in the middle part of the spectrum then extensively edited

309

Microsound in Composition

the sound le on a micro time scale to emphasize certain rhythms. I removed or

displaced many transients, for example, to decrackle the texture and articulate
more important transient bursts.
The sound world of Robotic is percussive. It begins with acoustic percussion
hits, many of which were pitch-shifted, ltered, or modulated to make variations on the original sounds recorded in the studio. As it progresses, more and
more synthetic tones appear, produced by pulsar synthesis, transient drawing,
and other methods.
Clang-Tint premiered at Kunitachi in 1994, and has since been performed
around the world.
Half-life
Half-life, composed in 1998 and 1999, explores the birth, replication, mutation,
and decay of sound particles. The composition divides into two parts, Sonal
atoms and Stratocumulus, with a total duration of six minutes. Work on the
piece began after experimentation with pulsar synthesis. (See chapter 4.).
Shortly after writing a pulsar synthesis program in the SuperCollider 1 language, I wrote a separate program for sound le granulation. In this granulation program, a constant-Q bandpass lter modies each sampled grain. The
frequency of the lter varies randomly within stipulated limits. (See the description of microltration in chapter 5.)
The source material of Half-life consisted of about forty-ve minutes of
stereo sound of two types. The rst type were raw pulsar trains with many
internal variations. The second consisted of granular ltrations of the rst.
Whereas the pulsar trains tended to be pointillist in texture, the ltrations were
more continuous and churning. The sound material in Sonal atoms is of the rst
type, while the material in Stratocumulus is of the second.
A period of dedicated editing followed the synthesis of the source material.
Much of the composition was assembled on the micro time scale, editing and
ltering individual particles. By lining up the particles in rapid succession, I
could induce the illusion of tone continuity. (See the description of particle
cloning synthesis in chapter 4.) Particle cloning shifted the perceived time scale
of a particle from the micro to the sound object level. I also introduced new
particles by transient drawing. (See the description of transient drawing in
chapter 4.) As a result, the tendency toward recycling and repetition, which is
still present in the work, was somewhat broken up. At the end of the day, many
particles in the piece were singularitiesunique events appearing once only.

310

Chapter 7

Figure 7.3 Concert hall of the Australian National Conservatory, Melbourne, prior to
the performance of the premiere of Half-life, May 1998. View from the stage. A loudspeaker is suspended at the center.

Half-life received its world premiere in May 1998 at the Next Wave Festival
in Melbourne, Australia, with sound projection over twenty-eight loudspeakers
(gure 7.3). The revised version premiered at the SuperCollider Night School at
the Center for New Music and Audio Technologies (CNMAT), University of
California, Berkeley in July 1999.
Tenth vortex and Eleventh vortex
On the night of 20 October 2000, I performed eleven ``sound cloud vortices.''
These consisted of real-time granulations of a single sound le: a train of
electronic impulses emitted by the PulsarGenerator program. Granulation
expanded the time base of the original by a factor of up to six, while also ltering each grain and scattering it to a random point in space. The swirling
combination of thousands of individual grains makes up the vortex.

311

Microsound in Composition

I chose the tenth and eleventh performances for further treatment. I cut the
Tenth vortex into four pieces, then nine more pieces, tuning and tweaking on a
micro time scale. The work proceeded rapidly. I linked the parts into the nal
version on Christmas Eve 2000. Eleventh vortex (2001) called for more nonlinearity in the macrostructure. I divided it into over eighty fragments, which
resulted in a much more complicated compositional puzzle, and a more idiosyncratic structure, alternating between coalescence and disintegration. Both
vortices premiered at Engine 27 in New York City in February 2001.
Microsound Techniques in Works by Various Composers
This section, which presents examples of works by various composers, cannot
hope to be exhaustive. The goal is to indicate how microsonic techniques are
being used by others.
Barry Truax
Several years after my article on the implementation of digital granular synthesis (Roads 1978a), the technique began to evoke interest for other composers. The Canadian Barry Truax developed the rst of several implementations
in 1986. His implementations are notable for their emphasis on real-time operation. Real-time synthesis is inevitably stream-oriented, and the musical aesthetic explored by Mr. Truax reects this orientation. Since the mid-1980s, he
has applied granular synthesis as a central technique in his oeuvre. His primary
emphasis is on the real-time granulation of sampled sounds, where he introduced many innovations. He has documented these in numerous articles (Truax
1986, 1987, 1988, 1990a, 1990b, 1991, 1992, 1994a, 1994b, 1995, 1996a, 1996b).
Truax was the rst composer to explore the gamut of eects between synchronic and asynchronic granular synthesis, which he employed eectively in a
series of compositions. In Riverrun (1986, Wergo WER 201750), from 5:40
to 6:50, he generated a series of synchronic ``steps'' consisting of overlapping
grains that simultaneously drift apart and lengthen, forming a melodic line.
Wings of Nike (1987, Cambridge Street Records CSR CD-9401, also Perspectives of New Music CD PNM 28) was the rst granular piece to use a sampled
sound as its source. The entire work is derived from two 170-ms phonemes,
magnied by the composer into a twelve-minute opus (Truax 1990b). Mixed
from an eight-track master, the work explores evolving streams of synchronic and asynchronic grains which become increasingly dense and merge into

312

Chapter 7

massive droning rivers of soundup to eight thousand grains per second by

the end of the work. Tongues of Angels (1988, Centrediscs CMC CD-4793)
features pulsating grains that widen in spectrum as their duration shrinks. Beauty
and the Beast (1989, Cambridge Street Records CSR-CD 9601) was the rst
composition to employ the process of real-time granulation of continuously
sampled sound, pioneered by the composer.
Much of Truax's output promotes the notion of soundscape, a sound portrait
of an environmental ambiance (Schafer 1977; Truax 1984). An early example of
soundscape music is Luc Ferrari's opus Presque rien, numero 1, composed between 1967 and 1970 (Robindore 1998). Here is Truax's description of soundscape composition:
The soundscape composition is a form of electroacoustic music, characterized by the presence of recognizable environmental sounds and contexts, the purpose being to invoke the
listener's associations, memories, and imagination related to the soundscape. At rst, the
simple exercise of `framing' environmental sound by taking it out of context, where often it
is ignored, and directing the listener's attention to it in a publication or public presentation,
meant that the compositional technique involved was minimal . . . This `neutral' use of the
material established one end of the continuum occupied by soundscape compositions,
namely those that are the closest to the original environment, or what might be called
`found compositions.' Other works use transformations of environmental sounds . . . with an
inevitable increase in the level of abstraction. However, the intent is always to reveal a
deeper level of signication inherent within the sound and to invoke the listener's semantic
associations without obliterating the sound's recognizability. (Truax 1996b)

It is in the transformational vein that Barry Truax worked in his soundscape

composition Pacic (1990, Cambridge Street Records CSR CD-9101). The
piece starts with four sound sources: ocean waves, boat horns, seagulls, and a
Dragon Dance recorded on the Chinese New Year. All sounds are heard at
their original pitch but are time-stretched through granulation, with up to
twelve stereo layers producing a chorus eect. Similarly, Pacic Fanfare (1996)
contains ten ``soundmarks'' recorded in the Vancouver area in the early 1970s
and the 1990s, reecting the changing soundscape of the city. The signals
appear both in their original state and granulated with resonance ltering and
time-stretching.
Horacio Vaggione
As early as the 1970s, Horacio Vaggione explored a musical aesthetic based on
the generation of a dense fabric of microsonic events. La Maquina de Cantar

313

Microsound in Composition

(1978) for example, applied an IBM computer to control a Moog synthesizer,

generating up to ten sounds per second in each of four voices (Bvdon 2000).
His output in the 1980s can be seen a bridge period. Examples from the 1980s
involving microsonic techniques and multiscale perspectives (using Music-N
computer languages for synthesis and transformation) are Thema (1985, Wergo
WER 2026-2), realized at IRCAM, and Tar (1987, Le Chant du Monde LCD
278046/47), realized at the Technische Universitat Berlin Elektronisches Studio,
also Kitab (1992, Centaur CRC 2255), for bass clarinet, piano, contrabass and
computer-processed sound.
As noted in chapter 5, Horacio Vaggione's 1995 composition Schall for tape,
commissioned by the Institut International de Musique Electroacoustique in
Bourges, is an outstanding example of the use of creative granulation. It represents a considerable renement of technique. The sound material of Schall
consists of tens of thousands of sound particles derived from sampled piano
notes.
The music is focused on a limited collection of objects of dierent sizes, which appear in
diverse perspectives, all along the process. The work plays essentially with contrasts between textures composed of multiple strata, as an expression of a concern with a detailed
articulation of sound objects at dierent time scales. (Vaggione 1999)

What makes Schall unique is its brilliant use of switching between dierent
time scales: from the microscopic up to the note-object level and down again
into the microscopic. Of course, the shorter the notes, the more broadband the
texture, as in the noisy section between 2:10 and 2:28, or the nal thirty seconds
of the work. Thus the interplay is not just between durations, but also between
pitch and noise.
Of the composition process, which involved interactive sound editing and
mixing software, the composer says:
Considering the handcrafted side, this is the way I worked on Schall (along with algorithmic generation and manipulation of sound materials): making a frame of 7 minutes and
30 seconds and lling it by ``replacing'' silence with objects, progressively enriching the
texture by adding here and there dierent instances (copies as well as transformations of
diverse order) of the same basic material. (Vaggione 1999)

In his studio, the composer demonstrated to me some of the micromontage

techniques used to make Schall. These involved arranging microsounds using a
sound mixing program with a graphical time-line interface. He would load a
catalog of pre-edited microsounds into the program's library then select items
and paste them onto a track at specic points on the timeline running from left

314

Chapter 7

to right across the screen. By pasting a single particle multiple times, it became
a sound entity of a higher temporal order. Each paste operation was like a
stroke of a brush in a painting, adding a touch more color over the blank space
of the canvas. In this case, the collection of microsounds in the library can be
thought of as a palette. Since the program allowed the user to zoom in or out in
time, the composer could paste and edit on dierent time scales. The program
oered multiple simultaneous tracks on which to paste, permitting a rich interplay of microevents.
With Nodal (1997), Vaggione has taken the materials used in Schall several
steps further, while also opening up his sound palette to a range of percussion
instruments. The identity of these instruments is not always clear, however,
since they articulate in tiny particles. The composition lasts 13:06, and divides
into three parts, with part I lasting until 5:46, and part II spanning 5:49 to 9:20.
The strong opening gesture establishes immediately the potential force of the
granular texture, and sets up a dramatic tension. Although the continuously
granulating texture that follows is often quiet in amplitude, one senses that the
oodgates could burst at any moment. This eect is enhanced by ``creaking''
sounds that give the impression of reins being strained.
Part II begins with a warm uttering texture that turns into a chaotic noise.
While the ear tracks this low-frequency rumbling, at 6:18 a distinct mid-high
crotale ``roll'' with a sharp resonance at 1600 Hz sweeps across. The overall
texture becomes unpredictably turgid and chaotic, until at 7:11 the composer
introduces an element of stasis: a rapidly repeating piano tone during which the
granulation background briey lets up. This leads to a section of almost tactile
noise, soft like a wet snowstorm. At 8:46 another wood tapping appears. Part II
ends on an incongruous major chord from what sounds like a toy piano. Part
III introduces a resonant tom-tom-like tone. The background texture is high
in frequency, sounding like rain on a thin roof. The density of the texture
gradually builds, as new bursts and resonances sweep into view. The texture
ebbs at 11:35, letting up until 12:09. The penultimate texture (a low-frequency
rumbling that also concludes Agon), is a long 39-second fade out. This texture
continues (at a low amplitude) for several seconds after the nal gesture of the
piece, a concluding three-event percussive tag ending.
The electroacoustic work Agon (1998) further elaborates the processes and
materials heard in Nodal. It opens with a continuously ``grinding'' band in the
range between 6 kHz and 16 kHz. The rate of the grinding modulation is in the
range of 10 Hz to 20 Hz. The continuity of the high-frequency band is broken
by various colored explosions. It is as if dierent percussive sounds are being

315

Microsound in Composition

Figure 7.4 A time-domain view of the nal gesture in Horacio Vaggione's Agon, a
triple-stroke ``tom-click-hiss.''

dropped into a gigantic granulator, to be instantaneously mulched into bits of

microsound.
Upon rst hearing, Agon appears to be a continuous stream of new material.
Repeated listening reveals that the work recycles material in an ecient manner. For example, the penultimate gesture of the worka swirling mid-low
frequency bandhas already been heard in the rst thirty-ve seconds. The
nal gesture of the work, a triple stroke ``tom-click-hiss,'' appears rst at 2:59
and again at 3:08 (gure 7.4). Two other distinctive timbres are a small metal
bell, and a tom-tom stroke. The bell is rst heard forty seconds into the piece,
resonating at about 750 Hz; at fty-nine seconds it shifts up to 1080 Hz (approximately an augmented fourth). The tom-tom is rst heard in a burst of
strokes at thirty-four seconds. Both the bell and the tom-tom reappear at many
points. A shimmering cymbal sound interweaves throughouta component of
the high-frequency band that streams through most of the piece. A piano tonecluster, rst heard at 2:01, signals the end of a quiet zone at 5:54, and marks a
turning point of the nale at 8:10.

316

Chapter 7

Figure 7.5 The amplitude envelope of the rst 61.8 seconds of Agon by Horaciio
Vaggione. The line marked T indicates the amplitude threshold between the foreground
peaks and the background granulations.

Microsound as a Foreground/Background Texture

A fascinating aspect of style in Schall, Nodal, and Agon, is the use of continuously granulating textures, composed of dense agglomerations of shortduration grains. These scintillating, sometimes crackling or frying textures serve
as a stationary element in the mesostructure of the pieces, holding the listener's
attention. By keeping these granulations low in amplitude, their background
role is evident. The composer sustains the low-level textures for ten to twenty
seconds or more at a time, keeping the listener engaged while he sets up the next
burst or explosive release (gure 7.5).
Musical Examples by Other Composers
This section oers capsule proles of pieces by other composers who have used
microsound synthesis and transformation. The techniques are now widespread,
and new compositions are appearing regularly. This survey, then, is little more
than a sampling of a larger body of work. The order of the descriptions is
roughly chronological.
Henri Pousseur's pioneering work Scambi (1957) applies serial principles to
``the nest particles of the musical-acoustical material,'' which in this case consisted of dense sequences of aleatoric pulses, resulting in white noise (Acousmatrix 4, compact disc BVHAAST 9010).

317

Microsound in Composition
As Michelangelo specied shape by chipping at his block of marble with a chisel, so
Pousseur specied crisp, clear, and pitched sounds by chipping at his block of white noise
with an electronic chisel called a lter. (Chadabe 1987)

The 1969 Chants de Maldoror by Reiner Riehn (Deutsche Grammaphon

137011) is a fascinating example of the use of microsonic materials in musical
synthesis. The sound material of this analog electronic work, composed at
the Institute of Sonology in Utrecht, consists of many thousands of ltered
impulses.
The second movement of Paul Lansky's As If (1982) for string trio and
computer-synthesized sounds (Centaur CRC 2110) features quasi-granular textures, which appear to be the result of echoes. The composer is also one of the
pioneers of micromontage, which he employed skillfully in a series of compositions including Idle Chatter (1985, Wergo 201050), in which the raw material
consists of vocal phoneme fragments of about 80 ms in duration scattered
metrically with a constantly shifting accent pattern. His subsequent pieces
just_more_idle-chatter (1987) and Notjustmoreidlechatter (1988) continue along
the same line, with strong emphases on articulation of meter and tonal harmonic progressions.
The technical environment of the 1970s and early 1980s made everything to
do with computer music rather laborious. Trevor Wishart's opus VOX 5, for
example, was conceived in 1979, but could not be realized until 1986, when the
technical climate was more benecent. VOX 5 is based on the phase vocoder
described in chapter 6. Although the utility of the phase vocoder was proven by
experiments in the mid-1970s (Portno 1976; Moorer 1978), it was not available to musicians until a decade later. This was rst possible with the CARL
software package from the University of California, San Diego (Moore 1982;
Dolson 1983, 1985, 1986). In VOX 5, Wishart creates a ``supervoice'' at center
stage, whose utterances metamorphose into natural events: crowds, bells, trees,
and less specic granular complexes. Another of Wishart's piece to combine
vocalisms with particle processing is Tongues of Fire, composed in 1995, about
which the composer has written:
I have newly invented many of the processes employed for this work. These include: Spectral Tracinga process in which a windowed Fourier analysis of the sound is made using
the phase vocoder; the sound is then reconstructed on a window-by-window basis using only
the N loudest components, the process itself is time-variable. Waveset Distortionthe time
domain signal is divided into elements at each even zero-crossing, and these elements are
manipulated in a variety of ways: for instance, each is repeated N times to produce a
strongly colored form of time-stretching which generates pitched ``beat streams'' within the

318

Chapter 7
sound; or N elements are replaced by just one of their number, extended in length to span
the entire set of N. The ``reworks'' transformation is produced by this procedure, etc.
Another process that is used, although less radically, is Sound Shredding, a technique related to Brassage. In this process a xed length of sound is cut into arbitrary length segments which are reordered at random, and this process is then repeated many times. The
progressive application of this process is heard in the voices to ``water''-like transformation
from 13 m 20 s to 14 m 40 s. (Wishart 1996)

Jean-Claude Risset made use of granular time expansion in Attracteurs

etranges. In this work, the clarinetist Michel Portal recorded a number of
scored musical phrases and sound eects. The composer transformed these
segments in various ways, including sequences speeded up or slowed down
using the Gabor grains (Risset 1992, 1998). In Sud (1984) and Elementa (1998),
both of which rely largely on eld recordings, Risset employed granular timeshuing. For Invisible for soprano and tape, he used the Gabor grains as implemented in the Sound Mutations program for MacOS (Arb and Delprat 1992).
Some of the earliest and most intensive use of phase vocoder transformations
is by the Santa Barbara composer JoAnn Kuchera-Morin. She has radically
altered the time, pitch, and spectrum structure of instrumental and vocal
sounds in a number of compositions including Dreampaths (1989), Cantata
(1989), Concerto for Clarinet and Clarinets (1991), and more recently Paleo
(2000) for double-bass and tape. A typical use would be to stretch the rst thirty
seconds of a sampled instrumental part into a three-minute layer. Another
technique is shrinking the duration of a melodic line from several seconds down
to a few hundred milliseconds, then expanding the duration of the shrunk line
back to the original length to emphasize the artifacts of the process. Using
Trevor Wishart's library of phase vocoder spectral operations, Kuchera-Morin
also made musical use of such advanced processes as morphing between two
sounds and spectral glissandi. Her work has involved three generations of phase
vocoders. The rst was Mark Dolson's pioneering implementation in the
UCSD CARL package with Wishart's library of transformations, the second
was the MixViews program developed by Douglas Scott at UCSB, the third a
variety of phase vocoders for MacOS.
Stephane Roy, in his 1992 composition Mimetismo (Empreintes Digitales
IMED 9630), explored a particulated world of microsounds based on the
plucked attacks of an acoustic guitar. The vocabulary of Mimetisimo relies
heavily on granulose manipulations. The work delights in the sensuous utter
of acoustic vibrations on the eardrum.

319

Microsound in Composition

Feuillages (1992) was composed by Philippe Schoeller for fourteen instruments and eight channels of electronic sound. The electronic part of this work
uses an ``acoustic pixel'' technique developed by the composer and his technical
assistant, R. G. Arroyo. The acoustic pixels (whose waveforms may be sampled
or synthesized) were generated in multiple streams (chemins straties), which
often overlap. The team then ltered the streams to create a ``liquid'' quality.
The Neapolitan composer Giancarlo Sica's En Sueno (Dreaming), realized in
1996, strives for a continuous transition from the comprehensible to the
incomprehensible. He achieves this through asynchronous granulation of the
Spanish word sueno sung by a tenor and repeated like an evocation. The granulation techniques were mixed with an additive synthesis texture that combined
thirty-two sine waves according to time-varying envelopes. The composition
was realized using the Csound language.
The Paris-based composer Gerard Pape has employed granular techniques in
combination with convolution in Makbenach (1997) for saxophone and tape.
In Makbenach, I worked with samples of various extended techniques for the saxophone,
developed and played by the saxophonist Daniel Kientzy. These were chained together to
make ``timbre paths.'' These timbre paths were composed as an alternative to isolated
``sound eects.'' That is, the paths involve a chaining together of isolated extended techniques to emphasize an overall timbral transformation, from simplicity to complexity,
purity to noise richness, harmonicity to inharmonicity, etc. (Pape 1998)

Pape used two programs to transform the timbre paths. First, he produced a
series of grains (using Cloud Generator) that followed a particular trajectory.
He used these as impulse responses to be convolved (using SoundHack) with
saxophone samples. The saxophone was transformed according to the path
of the grains. He also used Cloud Generator to create a series of evolving
granulations of the saxophone samples, establishing a new timbre path for
the saxophone, with the choppy rhythms of the grains.
Granular and phase vocoder techniques are central to certain compositions
of the German composer Ludger Brummer. His pieces scatter volcanic eruptions in a barely audible sea of sound. La cloche sans vallees (on Cybele CD
960.101) was realized in 1993 at the Stanford University studios in California.
The piece tends toward sparse, low-level textures, punctuated by ``explosive
interruptions'' (program notes). The Gates of H. (1993) expands a brief fragment of a folk melody sung by a female choir into an 18-minute composition.
After an explosive opening gesture, from about 1:10 to 2:07 the only sound is a
low-level reverberation which fades in and out of perception. A sharp burst
signals the start of the next section. The rest of the composition follows the

320

Chapter 7

Figure 7.6 The amplitude envelope of the 19-minute, 28-second composition Cri by
Ludger Brummer. The letters indicate crescendi sections. Note that the peak between C1
and C2 reects a momentary increase in low-mid-frequency energy (four pulsations). It
is not a crescendo.

same logic: sudden crescendo interjections follow long near-silent spaces.

The composer realized Dele! (1995, Cybele CD 960.102) at the studios of the
Academy of Arts in Berlin. It begins with samples of children shouting. These
were then time-stretched, pitch-shifted, and ltered using phase vocoders. Cri
(Edel 0014522TLR) employs the same materials and phase vocoder techniques.
As in his other compositions, dramatic gestures cause sounds to sweep into
view and then quickly disappear. For example, after a climax at 10:00 (E2 in
gure 7.6), the composition drops in amplitude by approximately 50 decibels
and drifts in and out of an inaudible zone for another minute-and-half. This
near-silence is broken by a dramatic ourish (F1 in gure 7.6).
The Franco-American composer Brigitte Robindore has also employed the
Cloud Generator program in a number of pieces. She has also developed techniques with microarcs on the UPIC system. Her Comme etrangers et voyageurs
sur la terre (1995) is a prime example of this latter technique. The mesostructure
consists of dense layers of sound masses that smoothly cross-fade into a constantly evolving ow. These are derived from samples of gamelan instruments
from the composer's own recordings, ltered with a variety of convolution
impulses. The piece combines these processed sounds with UPIC microarc
synthesis. The UPIC system generated multiple types of musical structures:

321

Microsound in Composition

grain clouds using fteen dierent envelopes simultaneously, repetitive polyrhythmic arc clusters based on sample extractions placed in infrasonic frequencies, and long sequences of superimposed sine waves that produce beating
eects in the range of 40 to 120 Hz. In these works, microsonic techniques serve
an aesthetic of emergent symbolism. In the composer's words:
Deeply embedded ideas slowly attain the threshold of perception and, at the right moment,
brim over into conscious thought. (Robindore 1996b)

In his 1999 doctoral dissertation, Manuel Rocha describes his three granular
synthesis compositions: Transiciones de Fase (1993), SL-9 (1994), and Moin
Mor (1995), which was realized with the Granular Synthesis Toolkit (Eckel, et
al. 1995). Moin Mor is based on an eighth century Irish poem and other, contemporary, poems in Gaelic. These poems mix with recordings made during the
composer's trip to Ireland in 1993, as well as sounds recorded by Italian journalist Antonio Grimaldi from Bloody Sunday (1972) in Londonderry. Moin
Mor begins with a granulated voice reciting the poems, using only consonant
sounds. The rest of the piece consists of multiple layers of granulated textures,
including phoneme fragments, over which the composer superimposes recognizable ``scenes'' from the Irish soundscape.
The media artist Kenneth Fields realized a compelling soundtrack for an
interactive science-education CD-ROM entitled Life in the Universe (1996).
Realized at CREATE, University of California, Santa Barbara, it featured
resonant wood-like grains and granulations of the computer synthesized voice
of the physicist Steven Hawking. According to the composer:
It was an appropriate metaphor, I thought, to use granular synthesis techniques for a
project having to do with the properties and history of the universe, particle and wave
physics, and the possibility of nding intelligent life in the universe beyond the Earth.
Hawkings' synthesized narration for the CD-ROM was recorded on tape and sent to me.
The piece has three parts, corresponding to three terrains we used as an organizational/
navigational strategy: the cosmological, biological, and mathematical terrains. Original
materials were derived both from sampled and synthesized (Csound) sources, then processed with Gerhard Behles's real-time Granular program Stampede II (Behles, Starke,
and Roble 1998). The music for the CD-ROM then, is a multiple path composition, dependent on the user's navigational choices. (Fields 1998)

Jon Christopher Nelson's granular opus They Wash Their Ambassadors in

Citrus and Fennel (Nelson 1996) was awarded a prize at the Bourges Festival in
1996. This piece uses granular textures as a background accompaniment to an
extended vocal performance by soprano Joan La Barbara.

322

Chapter 7

Damian Keller, a student of Barry Truax, developed a suite of programs for

producing granular sounds on models of such physical processes as bouncing,
breaking, scraping, and lling (Keller and Truax 1998). The results were made
available on an enhanced compact disc entitled touch `n' go (Keller 1999; earsay
productions es99002), which combined the text of these programs with original music compositions.
Natasha Barrett composed Little Animals in 1997. It appears on the 1998
Computer Music Journal compact disc. This composition, labeled ``acousmatic''
by the composer, employs ne granulations and particulations as a primary
strategy. According to the composer:
In Little Animals sound material rich in connotation opens the work, and then throughout
the last six minutes progressively fragments into pitch-textures. . . . The composition of
Little Animals began by recording source material already containing an allusive potential
(through mimetic approximation), which could be emphasized by straightforward editing
techniques (such as cutting, pasting and mixing). After establishing these allusions in the
composition, fragmentation techniques obscured the allusive aspect, revealing musical
qualities. The next transformation stage mainly involved ltering, spectral stretching and
fragmentation techniques. The fragmentation aspect is the most important as it serves to
concentrate the ear on the relationship between fragments, as opposed to the allusive content and the physical resonant properties of the original sound. This method allowed the
investigation of musical potentials without overt concern for the sounds' acoustic properties. (Barrett 1998)

In her doctoral dissertation Barrett described the work in detail, focusing on

the dierent listening strategies it tends to elicit (see also Barrett 1999).
The Italian composer Agostino Di Scipio has employed granular techniques
in a series of nine compositions, beginning with the tape work Punti di tempo
(1988) and extending to the ensemble and electronics piece Texture-Multiple
(2000). His approach emphasizes algorithmic control of microsound synthesis:
The task of microcompositional strategies can be described as one of letting global morphological properties of musical structure emerge from the local conditions in the sonic
matter . . . The fundamental point in such approaches lies in the fact that algorithms
implementing a model of sound material also implement ideas concerned with musical
form. (DiScipio 1994)

In his Sound & Fury, Di Scipio performs on a Kyma system (developed by

the Symbolic Sound company), eecting time expansion and frequency scaling
on an ensemble of acoustic instruments by using faders to control parameters
such as the time expansion, frequency, grain rate, and the spacing between
grains (Di Scipio 1997b).

323

Microsound in Composition

The developer of the Kyma system, the Illinois composer Carla Scaletti, used
it to time-stretch sounds from a telephone in an interactive installation called
Public Organ:
I used granular time stretching on the voices of installation participants who picked up a
telephone (with a high-quality microphone installed in place of the original telephone
microphone) and either left a ``voice mail message'' or were speaking with another installation participant over the net via CUSeeMe. I captured fragments of their voices and then
played them back a short time later time stretched. These time-stretched fragments were
stored on disk and became part of the sounds made by the installation in its other states,
long after the people had left. (Scaletti 1996)

Kim Cascone's Pulsar Studies 120 (1999) is an MP3 release (Internet:

fals.ch). The composer calls it ``a new exploratory collection of one-minute
studies composed using . . . pulsar synthesis.'' The sound material originated
from an early version of the PulsarGenerator program developed by Alberto de
Campo and me (see chapter 4).
Finally, Thom Blum's Five Haiku (2000) decomposes the soundscapes of San
Francisco into a swirling stream of sound particles, the sonic interpretation of
poems by the composer and Matsuo Basho (16441694).

Summary
[Of the Nobel prize winning physicist, J. J. Thomson, the discoverer of the electron] . . .
His early mathematical work was not outstandingly important. He was not skillful in the
execution of experiments. His talentone that is for both theorists and experimentalists
the most importantlay in knowing at every moment what was the next problem to be
attacked. (Weinberg 1983)

The synthesis and transformation of microsound has already aected the music
of our time. Viewing sound as a particulate substance opens up new approaches
to composition. Variations on the micro scale provoke changes in musical texture on higher time scales. Manipulations of particle density let coalescence and
evaporation function as event articulators, as they bring sound into and out of
being. Particle density also serves as the bridge between pulse and pitch, between rhythm and timbre. The plasticity of the sound, which is inherent in its
particle substrate, allows mutation to play an important compositional role,
since every sound object is a potential transformation.
Do we intend such operations to reduce compositional practice exclusively to
the level of microacoustical uctuations? No. They merely add this stratum to

324

Chapter 7

the rest of the known layers of composition, thereby enriching the eld. When
the microsonic layers interact with the higher layers, they tend to articulate
each other's specicity. We see this in the opposition of a long note underneath
a sparse cloud of impulses. But the situation is more complex, since we can
make any continuous sound evaporate at will. This introduces a fresh element
of dramatic tension to the unfolding of musical structure.
In the early years of digital synthesis, the musical possibilities inherent in
microsound were untested. This chapter shows that the situation is changing
and that important compositions employ these techniques. Utilizing microsound will require less justication in the future, it has already proved itself in
the most important arena, the domain of music composition.

Aesthetics of Composing with Microsound

Aesthetic Premises
The Philosophy of Organized Sound
Expansion of the Temporal Field
Illusions of Continuity and Simultaneity
A Multiscale Approach to Composition
Dierences Between Time Scales
Density, Opacity, Transparency
Stationary, Stochastic, and Intermittent Textures
Composition Processes on the Microsonic Level
Heterogenity and Uniqueness of Sound Materials
Aesthetic Oppositions
Formalism versus Intuitionism
Coherence versus Invention
Spontaneity versus Reection
Intervals versus Morphologies
Smoothness versus Roughness
Attraction versus Repulsion in the Time-Domain
Parameter Variation versus Strategy Variation
Simplicity versus Complexity in Microsound Synthesis
Code versus Grammar
Sensation versus Communication
Summary

326

Chapter 8
Aesthetics seems to thrive on controversy, even to demand it; on the conict, typically, of
new and old, of simplicity and complexity.
Edward Lippman (1992)
If styles and genres did not suer exhaustion, there would be only one style, one genre in
each art.
Jacques Barzun (1961)

An encounter with aesthetic philosophy is unavoidable in composition. To pick

up a gold-tipped fountain pen and inscribe a treble clef and key signature on
onionskin sta paper is to import a system of composition. It entails a selection from an enormous range of possible approaches to making music, a palette
of sounds, their combinations, a mode of performance, and even the audience. Every subsequent compositional decision further articulates the chosen
aesthetic.
No less a decision is taken in launching a program for granular synthesis,
which implies another context for creation, performance, reception, and evaluation of the resulting work. The dierences between these choices, which are
but two out of innumerable others, are dierences of aesthetic philosophy. An
aesthetic philosophy is nothing more than a collection of ideas and preferences
that inform the artist's decision-making. This collection does not determine
every artistic decision, but it guides the general direction. Nor is it static. Music
is in a constant state of change. The aesthetic sensibility of the creative artist is
continually evolving.
The purpose of this chapter is to try to identify the main aesthetic issues
raised by composition with microsonic materials. It is a worthwhile exercise for
a composer to formulate an aesthetic philosophy. Such reection forces one to
think through issues that may have been taken for granted. Reection upon
aesthetic ideas may lead to compositional ideas, and vice versa.
No matter how deliberately one composes, the aesthetic ramications of a
piece escape total control by the composer, who often has no power over the
theatrical presentation and the acoustical environment in which a work is performed. Even more important is the listener's mood, a narrow lter imposed on
the sonic sensation.
Before continuing, I should state that I do not mean to prescribe my aesthetics for others. My goal is simply to explain preoccupations that have guided
my own musical experiments.

327

Aesthetics of Composing with Microsound

Aesthetic Premises
Every doctrine of aesthetics, when put into practice, demands a particular mode of expressionin fact, a technique of its own. (Stravinsky 1936)

The present practice of electronic music composition rests on a foundation of

aesthetic premises. Some are old, others have emerged only recently. This section presents several of these premises.
The Philosophy of Organized Sound
Edgard Varese opened up a new path of exploration of musical sound when he
proposed, in the 1920s, a philosophy of ``organized sound.'' He was encouraged
by the early advances in music technology in electronic instruments such as
the Telharmonium, the Thereminovox, and the Ondes Martenot (Cahill 1897;
Rhea 1972, 1984; Chadabe 1997; Weidenaar 1989, 1995). For a time, Varese
also championed the experimental performances of the Futurist musicians,
builders of ``noise instruments.'' The Futurists wrote an emotional manifesto,
The Art of Noises, in which Luigi Russolo observed:
Musical sound is too limited in its variety of timbres. The most complicated orchestras can
be reduced to four or ve classes of instruments in dierent timbres of sound: bowed
instruments, brass, woodwinds, and percussion. Modern music ounders within this tiny
circle, vainly striving to create new varieties of timbre. We must break out of this limited
circle of sounds and conquer the innite variety of noise-sounds! (Russolo 1916)

The philosophy of organized sound extended the boundaries of accepted

musical material, and hence the scope of composition, to a wider range of
acoustic phenomena. (See also Cowell 1930; Cage 1937.) Creative musicians
sought beauty not only in the traditional, but also in the strange, in the formerly overlooked:
I insist upon the charm of combinations of noise instruments. But to appreciate this quality,
the most absolute silence is necessary in the concert hall. No one can imagine what charm
is attained with harmonic modulations and held chords produced, for example, by the blend
of low and medium howlers, low whistler, and hummer. What a marvelous contrast results
if a high crackler suddenly enters above this group to inect a theme, or a gurgler to hold
some notes or point up the rhythm! It is an eect that is completely unknown in orchestras,
since no orchestra but that of the noise instruments can produce this sensation of excited
and pulsing life, exalted through the intensity and rhythmic variety found in the combination of [the noise instruments]. (Russolo 1916)

328

Chapter 8

What critics dismissed as unmusical noise is now a potent element in the

composer's palette. Broad acceptance of this was slow in coming and Varese
encountered much rejection (Chou Wen-Chung 1966; Stuckenschmidt 1969).
After World War II, the musical avant-garde embraced him, though the general
public did not. Criticism and controversy surrounded his nal projects, Deserts
(1954, orchestra and tape), and Poeme electronique (1958, electronic tape).
Varese died in 1965.
Many post-WWII-generation composers sensed a great aesthetic crisis in
music. The twelve-tone ideas of Schoenberg and his followers had contributed
to the feeling that traditional methods of harmonic and rhythmic organization
were nearly exhausted:
The history of music and of musical thought is the story of articial systems, their inception, bloom, and decline, their absorption or replacement by other articial systems. . . .
Recent developments in the eld of musical composition have shown that the limited and
conditioned system of musical elements, considered musical material for several hundred
years, has now entered the administrative stage, where all possible permutations will no
longer possess any new meaning. (Brun 1970)

Simultaneous with this sense of crisis, new types of music were emerging out
of new musical materials. These included Pierre Schaeer's musique concrete,
and electronic music based on impulses, sine waves, noise generators, and
eventually, computer-generated sounds.
The aesthetic of organized sound places great emphasis on the initial stage of
compositionthe construction and selection of sound materials. This may involve synthesis, which often begins with microsounds, furnishing the elementary components used in the assembly of higher-level sound objects. Just as the
molecular properties of wood, thatch, mud, steel, and plastic determine the
architectural structures one can construct with them, so sonic microstructure
inevitably shapes the higher layers of musical structure.
The middle layers of musical structuremesostructurearise through interaction with the material. That is, to sculpt sonic material into gestures or
phrases involves mediation between the raw waveforms and the will of the
composer. This mediation is not always immediately successful, which is part of
the struggle of composition. If the initial result is unsatisfactory, the composer
has two choices. The rst is to develop new materials that will more easily t a
preconceived phrase mold. The second choice is to abandon the mold, which
means following the ``inner tensions''to use Kandinsky's phraseof the
sonic material (Kandinsky 1926). In this case, the material suggests its own
mesostructures. Later, the composer may intervene to reshape these structures
from the vantage point of another time scale.

329

Aesthetics of Composing with Microsound

These interrelationships between sound and structure conrm what musicians have known all along: material, organization, and transformation work
together to construct a musical code. It is in this context that a given sound
accrues meaning and beauty.
Expansion of the Temporal Field
Music theorists have long acknowledged a multiplicity of time scales in compositions. Today we can extend this awareness to the micro time scale. The call
for an expanded temporal eld was rst issued in the 1930s by composers such
as Henry Cowell and John Cage, who said:
In the future . . . the composer (or organizer of sound) will be faced not only with the entire
eld of sound but also with the entire eld of time. The ``frame'' or fraction of a second,
following established lm technique, will probably be the basic unit in the measurement of
time. No rhythm will be beyond the composer's reach. (Cage 1937)

By the 1950s, electronic devices had opened paths to the formerly inaccessible territories of microtime. In electronic music studios, one could assemble
complex sounds by splicing together fragments of magnetic tape. Composers
such as Stockhausen and Xenakis began to explore the temporal limits of
composition using these tape splicing techniques where, at a typical tape speed
of 38 cm/sec, a 1 cm fragment represented a time interval of less than 27 ms.
The analog signal generators of the 1950s let composers create for the rst
time sequences of impulses that could be transposed to dierent time scales by
means of tape speed manipulations. Designed for test purposes, the analog signal generators were not meant to be varied in time but favored a timeless wave
approach to sound. Their multiple rotary knobs and switches did not allow
the user to switch instantly from one group of settings to another. Because of
the weakness of their temporal controls, these devices imposed strict practical
limits, which, with assistance and a great deal of labor, one could work. (The
creation of Stockhausen's Kontakte comes to mind.)
By the 1970s, voltage-controlled analog synthesizers had become available,
manufactured by Moog, Arp, Buchla, and other small companies. Analog
synthesizers oered control through low-frequency oscillators, manual keyboards, and analog sequencers, but they could not provide for ne control at
the micro time scale. Neither was analog tape an ideal medium for organizing
microsonic compositions, owing to its inherent generation loss, linear access,
and razor-blade splicing. It was only with the dawn of computer synthesis and

330

Chapter 8

digital audio techniques that a micro approach to sound could be explored in

depth. (See chapter 2.)
Illusions of Continuity and Simultaneity
If events are absolutely smooth, without beginning or end, and even without modication or
perceptible internal roughness, time would nd itself abolished. It seems that the notions of
separation, . . . of dierence, of discontinuity, which are strongly interrelated, are prerequisites to the notion of anteriority. In order for anteriority to exist, it is necessary to be
able to distinguish entities, which would then make it possible to ``go'' from one to the
other. . . . Time, in a smooth continuum, is illegible. (Xenakis 1989)

Science has taken thousands of years to determine that the ne structure of

matter, space, and even time, is discontinuous and quantied (Hawking and
Penrose 1996). Human sensory organs set strict limits on our ability to perceive
the discontinuity beneath the apparently hard surface of all phenomena.
Discontinuous variations on the micro time scale of music melt the frozen
abstractions of traditional music theory such as pitch, instrumental timbre, and
dynamic marking. Even such sacred notions as tone continuity and simultaneity reveal themselves to be illusions. The micro time scale defrosts these
frozen categories into constantly evolving morphologies.
Continuity of tone is an auditory illusion. The mechanics of human hearing
smear streams of discrete events into an illusion of continuum. (See the section
on ``Perception of Microsound'' in chapter 1.) When we examine apparently
continuous tones under a microscope, we see gaps scattered throughout, like
the spaces between the threads of a blanket. These silent spaceson the order
of millisecondsare not perceived as temporally discrete events but as uctuations of a continuous tone. When two identical tones desynchronize on the
same time scale we perceive only a ``phase shifting'' eect, if anything at all.
A related illusion in music is simultaneity. As Xenakis (1989) observed, a
measurement as to whether two events occur simultaneously (or occupy the
same spatial position) depends entirely on the scale of observation. A detailed
analysis of onset times for supposedly simultaneous attacks in musical performance might reveal asynchronisms on the order of dozens of milliseconds.
A Multiscale Approach to Composition
A multiscale approach to composition allows for compositional intervention on
every time scale. The power of this approach comes from the fact that dierent

331

Aesthetics of Composing with Microsound

temporal zones interconnect. Operations on one time scale generate structures

which may be perceived on other time scales. This suggests the possibility of
working on the micro time scale in order to generate high-level musical structure. Iannis Xenakis predicted such an approach:
Suppose that each point of these [granular] clusters represents not only a pure frequency
. . . but an already present structure of elementary grains, ordered a priori. We believe that
in this way a sonority of a second, third, or higher order can be produced. (Xenakis 1960)

Interaction between the microtemporal scale and higher time scales is especially intriguing. To cite a simple example; a gradual change in particle durations results in timbre variations on a higher time scale. Certain signals cross
from one time scale to another, such as a descending glissando that crosses the
infrasonic threshold, turning from tone to rhythm.
For some composers, part of the attraction of composing with microsound is
the way it blurs the levels of musical structure:
The task of microcompositional strategies can be described as one of letting global morphological properties of musical structure emerge from the local conditions in the sonic
matter. (Di Scipio 1994)

In an approach that favors ``emergent properties'' or second-order sonorities,

sound objects emerge from a composer's manipulations of microstructural
processes (Di Scipio 1997a). This is also known as a bottom-up compositional
strategy, since the composition takes its shape from microsonic interactions.
This stands in contrast to a top-down strategy in which a composer lls in a
preplanned form (e.g., sonata).
The bottom-up strategy can be fascinating, partly because its results cannot
always be predicted in advance. On the other hand, why limit the scope of
compositional decisions to a single time scale? To navigate the widest possible
zone of creativity, the creative composer wants to oat freely across time scale
boundaries. To insert, delete, rearrange, or mold sounds on any time scale, this
is the multiscale approach to composition.
Dierences Between Time Scales
In the 1950s, a few serial composers tried to invent a system of composition that
could uniformly apply to any time scale. Their aim was for a kind of ideal
logical coherence. This did not take into account the nonlinearities of musical
perception. As chapter 2 indicates, a main lesson of Stockhausen's 1957 essay
``How time passes'' is precisely how awkward it can be to apply a proportional

332

Chapter 8

series developed for one time scale (e.g., pitch periods) to another time scale
(e.g., note durations). (Specically, it does not make much sense to transpose
the intervallic relations of the chromatic scale into the domain of note durations). Little music corresponds to a geometrically pure and symmetric hierarchical model. As Vaggione has stated:
The world is not self-similar. . . . Coincidences of scale are infrequent, and when one thinks
that one has found one, it is generally a kind of reduction, a willful construction. The ferns
imitated by fractal geometry do not constitute real models of ferns. In a real fern there are
innitely more accidents, irregularities and formal capricesin a wordsingularities
than the ossication furnished by the fractal model. (Vaggione 1996a)

Strictly hierarchical and symmetrical note relations are not necessarily perceived as such (Vaggione 1998). Gyorgy Ligeti (1971) pointed out the diculty
of organizing all time scales according to a unitary scheme, owing to a lack of
correlation with human perception. The musicologist Carl Dahlhaus (1970)
adopted a similar tack in his critique of serial pitch rule applied to the organization of microstructure.
Sound phenomena on one time scale may travel by transposition to another
time scale, but the voyage is not linear. Pertinent characteristics may or may
not be maintained. In other words, the perceptual properties of a given time
scale are not necessarily invariant across dilations and contractions. A melody
loses all sense of pitch, for example, when sped up or slowed down to extremes.
This inconsistency, of course, does not prevent us from applying such transpositions. It merely means that we must recognize that each time scale abides
by its own rules. A perfect hierarchy is a weak model for composition.
Density, Opacity, Transparency
The expansion of sonic possibilities adds new terms to the vocabulary of music.
We can now shape sonic matter in terms of its particle density and opacity.
Particle density has become a prime compositional parameter. Physics denes
density as the ratio of mass to volume. In music this translates to the ratio of
sound to silence. Through manipulations of density, processes such as coalescence (cloud formation), and evaporation (cloud disintegration) can occur in
sonic form. Opacity correlates to density. If the density of microsonic events is
sucient, the temporal dimension appears to cohere, and one perceives a continuous texture on the sound object level. Thus by controlling the density and
size of sound particles we have a handle on the quality of sonic opacity. Coalescence takes place when particle density increases to the point that tone

333

Aesthetics of Composing with Microsound

Figure 8.1 A hole in a broadband sound, sculpted by a sonogram lter. We can carve
time-frequency gaps on dierent time scales.

continuity takes hold. An opaque sound tends to block out other sounds that
cross into its time-frequency zone.
Going in the opposite direction, we can cause a sound to evaporate by
reducing its particle density. A sparse sound cloud is transparent, since we can
easily hear other sounds through it. A diaphanous cloud only partially obscures
other sounds, perhaps only in certain spectral regions. For example, by means
of sonogram ltering we can create transparent holes in the spectrum of a
sound (gure 8.1) which might provide a window onto another layer of sound
beneath.
Stationary, Stochastic, and Intermittent Textures
Many complex musical textures resemble what statistics calls stationary processes. A stationary process exhibits no trend. The texture has a xed mean
value and uctuates around the mean with a constant variance. A stationary

334

Chapter 8

process is not necessarily static in time, but its variations remain within certain
limits, and are therefore predictable. We see these characteristics in many
sound-mass textures created with particle synthesis. Consider a dense cloud of
grains scattered over a broad zone of frequencies. It scintillates while never
evolving, and is therefore a stationary texture.
Stationary textures are fertile material for composition. One can place them
at low amplitude in the background layer, where they lend depth to the musical
landscape. Positioned in the foreground, their constant presence introduces
dramatic tension and sets up an expectation of change. The ear notices any
change as a deviation from the stationary.
Changes in texture appear as slow trends or sudden intermittencies. To impose a trend is to gradually change the texture. This may take place over time
periods associated with the meso time scale (i.e., many seconds). A trend converts a stationary texture into a weighted stochastic texture. One can introduce a
trend by opening or closing the bandwidth of a cloud, by altering its center
frequency, by ltering, or by any other perceptible time-varying operation.
Sudden changes create intermittent texture. The intermittencies break up
the stationary texture by injecting loud particles or silent micro-intervals. This
latter techniquecomposing with silenceremains largely unexplored, but can
be eective. The idea is to begin with a multichannel stationary texture and
introduce silent intervals into it, working like a sculptor, carving rhythmic and
spatial patterns by subtraction.
Composition Processes on the Microsonic Level
Interactive digital sound editing originated in the 1970s (Moorer 1977a, 1977b),
but did not become widely available until the 1990s. It seems so taken for
granted today, I am not sure that musicians recognize its profound impact on
the art of electronic music composition. The possibility of editing sound on
any time scale has opened up a vast range of transformations. For example,
through selective gain adjustment it is possible to magnify a tiny subaudio
uctuation into an intense microsonic event (see the description of transient
drawing, chapter 4). We can shape the event through microsurgery or ltering
on a micro time scale (chapter 5). By replicating the event on the micro time
scale, it can be transformed into a pitched sound object (see the description of
particle cloning synthesis in chapter 4). Through time-stretching it can be
magnied into a long, slowly unfolding texture (pitch-time changing chapter 5,
and sound transformation with the phase vocoder, chapter 6). Then through

335

Aesthetics of Composing with Microsound

granulation we can decompose it once again into small particles, ``from dust to
dust'' (granulation, chapter 5).
Such manipulations open up a catalog of new compositional processes:
1 Variations (contrasts, increases, and decreases) of particle density
1 Coalescence (cloud formation) and evaporation (cloud disintegration)
1 Time stretching to extend a microstructure into a large-scale event
1 Time shrinking large events into microsounds
1 Hierarchical variations of the same event structure on multiple time scales
1 Lamination of a cloud through multiple layers with microtemporal delays
1 Particle spatialization (scattering particles in space)
1 Granular reverberation
1 Precise polymetric pulsations in space, created by superimposing multiple
metrical streams
1 Multiple formant streams, each with its own frequency and spatial trajectory
1 Spectrum evolution via manipulation of particle envelopes
1 Microsurgery on the Gabor matrix to extract the chaotic, harmonic, loudest,
softest, or other selected particles within a sound and reassemble it with
alterations
Such operations change the practice of composition, and mandate a rethinking
of compositional strategy and architecture. This cultural process has only just
begun.
Heterogenity and Uniqueness of Sound Materials
In the 1950s, certain composers began to turn their attention toward the composition of sound material itself. In eect, they extended what had always been
true at the phrase level down to the sound object level. Just as every phrase and
macro form can be unique, each sound event can have an individual morphology. This creates a greater degree of diversityof heterogeneity in sound materialwithout necessarily losing continuity to other objects. Chapter 3 showed
how we can extend the concept of heterogeneity even further, down to the level
of microsound, where each sound particle may be unique. The microstructure
of any sound can be decomposed and rearranged, turning it into a unique
sound object.

336

Chapter 8

Simultaneous with this shift in musical thinking, certain researchers have

shifted their focus from the analysis of periodicities to the analysis of singularities (Arneodo et al. 1995). They have turned from the analysis of continuous
signals to the analysis of intermittencies, that is, from stationary and homogeneous emissions to nonstationary and nonhomogeneous distributions of energy.
The variety of sound opened up by electronic music comes at a high price:
the loss of note homogeneity and with it, the foundation of traditional abstract
musical language. To adopt the universe of heterogeneous sound objects is to
be cast into a strange new land without conventional language. The terrain of
this land is nonhomogeneous, pocked by fractured disjunctions (intermittencies)
and nonlinear transitions from one perceived state to another. A simple linear
change may trigger such a percept, such as a pulse train whose frequency is
smoothly increasing from 5 to 50 Hz. In the midst of this change, listeners
perceive a series of rhythmic pulsations passing through an ambiguous zone
between rhythm and pitch into a continuous tone. An asynchronous cloud of
grains cross the threshold (from discontinuity to continuity) as the density
increases from ve to one hundred grains per second. Many other examples of
nonlinear thresholds could be cited.

Aesthetic Oppositions
It seems inevitable that we seek to dene and understand phenomena by positing their opposite. High cannot be understood without the concept of low, and
so with near and far, big and small, etc. A given aesthetic tendency can be seen
as confronting its opposite. The question is whether such a simplication can
lead to a clarication. This section explores certain aesthetic oppositions raised
in composing with microsound.
Formalism versus Intuitionism
In composing with microsound, we face an ancient conict: formalism versus
intuitionism. Formal models of process are natural to musical thinking. As we
listen, part of us drinks in the sensual experience of sound, while another part is
setting up expectationshypotheses of musical process. To the envy of the
other arts, notation and logical planning have been part of music-making for
centuries. As Schillinger (1946) demonstrated, we can make a music generator

337

Aesthetics of Composing with Microsound

out of virtually any mathematical formula. Lejaren Hiller's pioneering experiments with automated composition in the 1950s proved that the computer
could model arbitrary formal procedures (Hiller and Isaacson 1959). Computer
programs sped up the time-consuming labor associated with systematic composition. This led to a surge of interest in applying mathematical procedures to
composition (Hiller 1970).
Since the start of music notation, it has been possible to manipulate musical
materials as symbols on paper, separated from the production of sound in time.
Herein lies a fundamental dichotomy. Because formal symbols can be organized abstractly, such manipulations have been closely identied with the organization of sound material. Music, however, is more than an abstract formal
discipline. It must eventually be rendered into sound and heard by human
beings. Thus it remains rooted in acoustical physics, auditory perception, and
psychology.
One cannot escape formal control when working with a computer. Every
gesture translates into an intervention with a formal system. This system is
encoded in the logic of a programming language and is executed according to
the algebra of the machine hardware. The question is at what level of musical
structure do such formalisms operate? The pianist practicing on a digital piano
is interacting with a computer music system. She is not concerned that her
performance is triggering a urry of memory accesses and data transfers. The
familiarity of the keyboard and the sampled piano sounds makes the interaction seem direct and natural. This is a great illusion, however. With a change of
formal logic, the same equipment that produces the piano tones could just as
well synthesize granular clouds, as we saw with the Creatovox (chapter 5).
Applied at dierent strata of compositional organization, formal algorithms
can be a powerful means of invention. An algorithm for spawning sound particles can handle enormous detail in a fraction of a second. Other algorithms
can iterate through a collection of variations quickly, oering the composer a
wide range of selections from which to choose. Interactive performance systems
try to balance preprogrammed automation with spontaneous decisions.
While formal algorithms enable interaction with a machine, formalism in
composition means imposing constraints on one's self. The formalist composer
follows a logical system from beginning to end. This logic exists only in an ideal
conceptual plan. The plan must ultimately be translated into the real world of
acoustics, psychoacoustics, and emotional response. It is in this translation that
the game is often lost.

338

Chapter 8

Coherence versus Invention

Coherence must bear some relation to the listener's subconscious perspective. But is this
its only function? Has it not another of bringing outer or new things into wider coherence? (Ives 1962)

In academic theory, formal coherence is one of the most vaunted characteristics

of musical composition. In general, coherence signies ``logical integration and
consistency.'' This quality is not always easy to measure in practice. In its
most obvious form, coherence manifests itself as a limitation in the choice of
compositional materials and a consistency in the operations applied to those
materials.
One can easily place the organization of microsound under the regime of a
formal system. In this case, the operations that produce the composition ensure
that it always remains within the boundaries of the formal rules. Such an
approach makes for a tidy package, free from anomalies and logical inconsistencies. The compositions it produces can be proven to be ``coherent'' in a
formal sense, even if they are dull or incomprehensible.
The problem here, as we have stated before, is that music is not a purely
logical system. Rigor is not synonymous with perceived musical coherence.
Music is rooted in acoustics, auditory perception, and psychology. Musical coherence seems to be a poorly understood psychological category. It is one of
those ubiquitous terms in aesthetic discourse that everyone uses subjectively
and no one has ever studied from a scientic viewpoint.
As Horacio Vaggione (1997) observed, by convention, a ``rule'' must necessarily be followed many times. But the artist can invoke a rule only once! Thus
we might focus our attention on other criteria in the way that we compose
with microsound. Inventiveness, I would suggest, is at least as important as
coherence.
Spontaneity versus Reection
We nd in electronic music new examples of a venerable opposition in musicmaking, pitting the immediate spontaneity of improvisation in performance
against the careful, reective process of studio-based composition. This confrontation is particularly sharp in the case of real-time systems that spawn a
constant swarm of sound particles. To control this ow in such a way as to
make interesting musical gestures is not easy. The musician's interface can
either help or hinder this process.

339

Aesthetics of Composing with Microsound

Barry Truax's granular synthesis programs GSX and GSAMX (Truax 1988)
incorporated several types of controls:
low-level grain parameterscenter frequency, frequency range, average duration, duration range, delay (density)
presetsgroups of stored grain parameter settings
rampspatterns of change in grain parameters, stored in a ramp le
tendency masksgraphic control shapes that are translated into ramps and
presets, stored in a tendency mask le
The composer could override any of these stored parameters in performance,
intermingling planned functions with spontaneous gestures.
Our Creatovox instrument (chapter 5) takes another approach to the problem of particle synthesis in real time. Designed as a solo instrument for virtuoso
performance, it is played using a traditional keyboard, with additional joystick,
fader, and foot pedal controllers. In the Creatovox, each keystroke spawns a
cloud of grains, whose parameters can be varied with the real-time controllers.
PulsarGenerator (chapter 4) was not intended for the concert-hall, although
we anticipated that it would be used in this way, because it could be operated in
real time. We wanted a program which would allow improvisation as a fast way
to explore the wide range of pulsar textures, but which also allowed for careful
planning through control by envelopes and stored presets.
Despite the attractions of real-time music-making, the studio environment
is the ultimate choice for the musician who seeks the maximum in creative
freedom:
1 The possibility of editing allows any previous decision to be revised or
retracted in the light of reection.
1 Rehearsal of all gestures permits renement.
1 In contrast to real-time improvisation, where the focus tends to be local in scope,
studio decision-making can take into account the entire range of time scales.
1 An arbitrary number of independent musical threads can be superimposed
carefully via mixing.
1 The sound structure can be monitored and manipulated on a particle-byparticle basis, which is impossible in real time.
A potential hazard in studio work is over-production. An over-elaborate
montage may result in a stilted and contrived product.

340

Chapter 8

Intervals versus Morphologies

Atomism compels us to give up the idea of sharply and statically dening bounding surfaces
of solid bodies. (Einstein 1952)

Linked to a wave-oriented view of sound is a classical aestheticdating back to

the Greeksthat assigns great value to works of art that conform to certain
numerical proportions and intervals. This aesthetic imprints itself throughout
the history of music, particularly in the domain of pitch relations. It is also
implicit in the treatment of metrical rhythms, with its scale of durational values
based on duple and triple divisions of the beat.
Intertwined with intervallic thought is the notion of scales. As Karlheinz
Stockhausen indicated (1957, 1962), any continuous musical parameter (spatial
position, lter setting, etc.) can be subdivided into an arbitrary scale and then
manipulated in terms of intervallic relations. The twentieth century saw the
introduction of serial, spectral, and minimalist aesthetic theories, all of which
were intervallic. The main dierences between them concern which intervals
and which scales are most important.
Acoustic and perceptual reality stand in contrast to the simplications of
intervallic thought. The momentary frequency of most acoustic instruments is
constantly changing. Noise is ubiquitous. Dierence thresholds limit all aspects
of perception. Masking and other nonlinear eects complicate perception.
Training and mood strongly inuence musical hearing.
To think in terms of microsonic materials and procedures is to shift the aesthetic focus away from sharply dened intervals toward curvilinear and fuzzy
morphologies. Just as it has become possible to sculpt habitats from berglass
foam, the owing structures that we can create with microsound do not necessarily resemble the usual angular forms of musical architecture. On the contrary, they tend toward droplike, liquid, or cloudlike structures.
Sound particles dissolve the solid notes into more supple materials which
cannot always be measured in terms of denite intervals. As a result, sound
objects may have ``fuzzy edges,'' that is, ambiguous pitch and indenite starting
and ending times (owing to evaporation, coalescence, and mutation). Microvariations melt the frozen abstractions of traditional music theory such as continuous tone, pitch, instrumental timbre, and dynamic marking, reducing them
to a constantly evolving stream of particle morphologies. Intervals may emerge,
but they are not an indispensable grid, there is instead an interplay between
intervallic and nonintervallic material.

341

Aesthetics of Composing with Microsound

Within these owing structures, the quality of particle densitywhich determines the transparency of the materialtakes on prime importance. An increase in density induces fusion. It lifts a cloud into the foreground, while a
decrease in density causes evaporation, dissolving a continuous sound band
into a pointillist rhythm or vaporous background texture. Keeping density
constant, a change in the characteristics of the particles themselves induces
mutation, an open-ended transformation.
Smoothness versus Roughness
The shapes of classical geometry are lines and planes, circles and spheres, triangles and
cones. They inspired a powerful philosophy of Platonic harmony. . . . [But] clouds are not
spheres. . . . Mountains are not cones. Lightning does not travel in a straight line. The new
geometry models a universe that is rough, not rounded, scabrous, not smooth. It is the
geometry of the pitted, pocked, and broken up, the twisted, tangled, and intertwined. . . .
The pits and tangles are more than blemishes distorting the classical shapes of Euclidean
geometry. They are often the keys to the essence of the thing. (Gleick 1988)

Microsonic synthesis techniques contain a dual potential. On the one hand,

they can create smooth and pitch-stable continua, characteristic of the formants
of the voice and resonant instruments. On the other hand they can create
intermittent particles and nonstationary textures, which in the extreme tend
toward chaotic noise bands.
The determinants of pitched continua are stable waveforms, round envelopes, and long particle durations. In contrast, the determinants of noisy signals
are irregular waveforms, jagged envelopes, and brief particle durations. The
opposition between smooth and rough textures can serve as an element of tension in composition, akin to the tension between consonance and dissonance,
with transitions between these two extremes acting as a bridge.
Attraction versus Repulsion in the Time-Domain
Astronomers from the Paris Observatory have detected the densest mass of galaxies that is
visible in the southern hemisphere, which may constitute the central element of the Great
Attractor. Predicted in 1987, the Great Attractor would be an enormous concentration of
galaxies that attracts other galaxies in its environmentincluding our ownat speeds on
the order of 500 to 600 km per second. (Associated Press France 1997)

The universal principle of attraction and repulsion governed the primal cosmological explosion of the Big Bang as well as the inner structure of atomic

342

Chapter 8

particles. It manifests itself in physical biology as the experience of pleasure and

pain, and in the psychological experience of love and hate, of lust and disgust.
It rules over individual human relationships as much as relationships between
tribes and cultures.
The principle of attraction and repulsion also plays a role in music. For example, Igor Stravinsky used attraction as a means of organizing the time
structure of a composition:
Composing leads to a search for the . . . center upon which the series of sounds . . . should
converge. Thus if a center is given, I shall have to nd a combination that converges on it.
If, on the other hand, an as yet unoriented combination has been found, I shall have to nd
a center towards which it will lead. (Stravinsky 1947)

Edgard Varese thought that it might be possible to adapt the principle of repulsion as an organizing principle:
When new instruments will allow me to write music as I conceive it, taking the place of the
linear counterpoint, the movement of sound-masses, or shifting planes, will be clearly perceived. When these sound-masses collide, the phenomena of penetration or repulsion will
seem to occur. (Varese 1971)

Temporal attraction takes three forms: attraction to a point, attraction to a

pattern, and attraction to a meter.
Attraction to a point refers to when numerous particles gravitate toward a
specic time point, and the clustering results in a climactic explosion. Its
conceptual opposite is emptiness or silence, or the repulsion of sound objects
away from a central point. I have applied these concepts in my piece Clangtint, where points of attraction control the density of sonic events (see chapter 7).
Attraction to a pattern refers to a strong tendency toward reoccurrence of a
given rhythmic motive. The isorhythms of ancient music exemplify this phenomenon. Repulsion from a pattern refers to the absence or avoidance of regularity in motivic guration.
Attraction to a meter means alignment to a regular pulsation. A strong
metric beat attracts a metric response. It is easy to synchronize multiple layers
on top of a regular pulse. The opposite of metric attraction is metric repulsion,
found in the rich realm of ametric (or aleatoric) rhythms. It is dicult to overdub synchronously on top of an ametric rhythm. An ametric rhythm is not the
same as syncopation; syncopation reinforces meter by emphasizing various of
its subdivisions.

343

Aesthetics of Composing with Microsound

Parameter Variation versus Strategy Variation

In electronic music, every technique of synthesis is controlled by a number of
parameters. This invites a process of composition based on parameter variation. Consider the technique of frequency modulation synthesis (Chowning
1973). The timbre of many computer music pieces of the 1970s and 1980s relied
on variations of essentially two parameters: the index of modulation and the
carrier-to-modulator ratio. In granular synthesis, variations in grain spacing
and grain density lead to time-stretching and time-shrinking eects, while variations in grain duration lead to spectrum transformations. A prime example of
parameter variation is Barry Truax's composition The Wings of Nike (Perspectives of New Music 28 compact disc, 1990). Here two phonemes, each about
170 milliseconds in length, iterate over a 12-minute composition, a repetitionexpansion by a factor of four thousand.
As a compositional strategy employed to the exclusion of others, parameter
variation reects an obsession with consistency. It can take the form of a preoccupation with deriving every sound from the previous one, or from a source
sample. If this is carried out only through variation of a single algorithm, it
may lead to a monochromatic timbral palette. It may also result in a restricted
range of gestures, since no morphological developments can occur that are not
derived from parameter variation.
Alternatively, the compositional strategy can itself be the subject of variations. Even a simple change, such as switching to a dierent synthesis technique, alters the parameters. Another strategic change is a step up or down to a
dierent level of structure. Juxtaposition refreshes the brain, breaking the cycle
of closed permutations and combinations.
Simplicity versus Complexity in Microsound Synthesis
The synthesis instruments presented in chapters 3 and 4 are simple in structure,
reecting my own approach to the synthesis of microsound. Certain basic
sounds, like the magical sine wave, can be made expressive with only a touch
of vibrato and tremolo, and perhaps a dash of reverberated ambiance. Most
other interesting sounds, however, are more complicated in their time-varying
behavior.
From the standpoint of the composer who is also a programmer, the question
is whether to embed such complex behavior within a synthesis instrument, or
whether to separate the behavior from the instrument. My tendency is towards

344

Chapter 8

the latter. Rather than designing a latter-day Wurlitzer organ controlled by

dozens of parameters, I prefer to build a library of small distinct instruments,
each with its own articulators and modes of performance. In this approach, the
score serves as the primary point of control. Therefore, the score is the source of
synthesis complexity. As in the traditional practice of orchestration, it is by
selecting and combining dierent instruments that one achieves a broad palette
of timbral colors. The score interweaves these colors in detail. In electronic
music, of course, the score does not necessarily take the form of a ve-line sta.
It can be a collection of sound events and envelopes, as in the note lists of the
Music N languages, or the graphical regions of a sound mixing program (e.g.,
Digidesign's Pro Tools).
When particles assemble into sound objects, it is their combination en masse
that forms the morphology of the sound. By controlling this combination from
a high level, we shape the sound's evolution. High-level controls imply the existence of algorithms that can interpret a composer's directives, translating
them into, potentially, thousands of particle specications. An early example of
such an algorithm is my 1975 grain generator program PLFKLANG, described
in chapter 7. This trend is now re-emerging in the form of high-level generators
of microsonic behavior. A recent example is the program Cmask (Bartetzki
1997a, 1997b). This is a stochastic event generator which works in conjunction
with the Csound synthesis language. The program reads a composer-specied
parameter le and generates a score le that Csound can immediately read.
For every parameter of an event, Cmask selects a random value from a timevarying tendency mask. CMask also provides other methods of parameter
generation, including cyclic lists, oscillators, polygons, and random walks. A
dierent method may generate each parameter of an event.
James McCartney's SuperCollider language provides another set of highlevel event generators, called Spawn, Tspawn, OverlapTexture, and XFadeTexture which generate a variety of behaviors that can be edited and adjusted
by the user. For example, OverlapTexture creates a series of overlapped sounds,
where the sounds are generated by a user supplied synthesis instrument. By
adjusting the parameters of the OverlapTexture, and randomizing some of
them within specic limits, one can create a wide variety of ambient textures.
To conclude, a synthesis technique is a means to an end. In my work, synthesis is the starting point for sound design, which is itself only the beginning of
composition. I inevitably edit (cut and paste) and alter raw sound material with
a variety of transformations. This editing involves trial-and-error testing and

345

Aesthetics of Composing with Microsound

renement. Because of this, it would be impossible to bundle into a synthesis

algorithm.
Code versus Grammar
If played and heard often enough, every musical gesture is prone to be interpreted by
musicians and listeners, as a gesture of musical speech. As the gesture becomes familiar,
and thus recognized by society, the composed structure, in which the context generates the
meaning of its components, will be misunderstood, instead, as one in which the components
give meaning to their context. (Brun 1983)

Traditional western music is characterized by a familiaryet fundamentally

incomplete and ambiguousmusical grammar. This grammar describes the
materials employed in pieces (pitches and timbres) as well as many aspects of
their organization. In contrast, in new music, particularly pieces that employ
microsonic synthesis, the audience does not necessarily know the grammar in
advance. Here, it is the task of each new piece to establish its own code, with
perhaps only a few symbols inherited from the past. The term ``code'' distinguishes the provisional nature of the communication protocol, as opposed to
a more established language with its grammar.
When we can recognize a familiar unit of musical structure, we might say
that it is a cliche. The popular understanding of ``cliche'' is that it is equivalent
to ``overly familiar.'' Yet frequency of exposure seems to have little to do with
the issue; lovers of Mozart never tire of their hero's recognizable sound. Cliche
is a pejorative in colloquial English, where it is a synonym for ``hackneyed.'' Its
meaning is more neutral in its original French, where it may also refer to a
templatean original from which many instances can be derived.
The danger of cliches, in the narrow, popular sense, is omnipresent. The naive presentation of trite formulas (as if they retained any freshness) creates a
negative impression. The denition of what is fresh and what is trite is subjective, however, and varies with ckle artistic fashion. Music history shows
a constant struggle between progressive and reactionary trends, in which the
ascendence of each alternates every few years. As progressive movements stop
evolving, they become reactionary, while trends that formerly seemed hackneyed, seem suddenly fresh or even daring in comparison. (A classic example
is the rise of simple tonal music in the 1970s and 1980s after the peak of
serialism.)
In articial intelligence research, the term ``cliche'' has been used with another meaning, and it is this sense that I would like to emphasize here. In this

346

Chapter 8

context, a cliche is a unit of knowledge that corresponds to a formal structure

such as a semantic network (Winston 1984). Semantic networks oer a formal
structure for capturing the essential properties of a known class of objects or
situations, enabling programs to recognize instances of them. If, for example,
we dene a table as an object with a at surface supported by one or more legs,
a program should be able to classify all such objects, in many shapes and sizes,
as instances of a single cliche. David Cope (1996) uses the term ``signature'' in a
manner similar to my use of ``cliche.''
The notion of cliche or signature in music refers to memorable units of musical structure, and implies a process of recognition on the part of the listener. It
is important that the cliche be memorable. Given enough of the cliche we can
``ll in the blanks'' and substitute missing pieces. As with the table example, we
can spot a cliche in many guises.
Naturally, to spot a musical cliche requires sensitivity and erudition. Neophytes cannot spot cliches as readily as experts and some cliches are stronger or
easier to recognize than others. But the ability to recall entire pieces of music
from fragments of as few as two or three notesmore or less instantaneously
reminds us of how human perception is attuned to cliche (Buser and Imbert
1992). Perhaps the notion of cliche can be applied in a practical agenda to make
computers more musically intelligent.
Through the accumulation of cliches over time, understandable codes inevitably emerge from the magma of artistic creation. Strange musical landscapes
become more familiar. The vintage sonorities of early electronic music, for
example, are no longer novel. The basic textures of granular synthesis are becoming familiar. The practice of composition with microsound must eventually
shift from the creation of novel sonic eects to the construction of compelling
new musical architectures.
Sensation versus Communication
Within a fraction of a second after the eyes, nose, ears, tongue or skin is stimulated, one
knows the object is familiar and whether it is desirable or dangerous. How does such
recognition, which psychologists call preattentive perception, happen so accurately and
quickly, even when the stimuli are complex and the context in which they arise varies?
(Freeman 1991)

Sound waves speak directly to our senses. They can be likened to the immediate
perception of touch, if touch could penetrate to the inner ear. The experience of

347

Aesthetics of Composing with Microsound

music is a cognitive reaction to a perceptual reaction. It reaches directly toward

emotions and associations. Intellectualization is a side eect.
Traditional musical languages abide by familiar grammars. This familiarity
acts as a framework for setting up small surprises. In creative electronic music,
which attempts to extend musical expression itself, the goal is often the opposite. Since so much is new, little is expected. The surprise comes in nding
familiarity.
In this music, the role of the composer is to create a pattern of acoustic sensations in the form of a code that organizes them into a meaningful structure.
(I will not attempt to dene ``meaningful'' here.) The intellectual challenges and
emotions experienced by the composer in creating this structure may be very
profound and intense (or not). In any case, they are independent of those experienced by the listener. The composer cannot hope to account for the mindset
carried into the concert hall by the listener. Acoustic sensations are inevitably
ltered by the listener through the extremely narrow sieve of subjective mood
and personality. These interpretations trigger a reverberant halo of emotions
and reections unique to each person.
The ideal of musical communication would imply a direct transmission of
emotional and intellectual experience from composer to listener. This is probably quite rare. Is direct transmission the point? The point seems to be to stimulate the listener with organized sensations. The composer takes the listener on
a fantastic voyage. Let each person make up their own mind about what they
experience.

Summary
Every work of art aims at showing us life and things as they are in truth, but cannot be
directly discerned by everyone through the mist of subjective and objective contingencies.
Art takes away the mist. (Schopenhauer 1819)
Art, and above all, music, has a fundamental function . . . It must aim . . . toward a total
exaltation in which the individual mingles, losing consciousness in a truth immediate, rare,
enormous, and perfect. If a work of art succeeds in this undertaking, even for a single moment, it attains its goal. (Xenakis 1992)

Art music is designed according to an aesthetic vision. It is formed by a blend

of pure sensation and logical organization. The joy of composition derives
from the free interplay between goal-driven intellectual and emotionally driven

348

Chapter 8

intuitive modes of working. The act of composition cannot avoid expressing

philosophical attitudes, even if these are vague (ambiguous) or even contradictory (ironic).
In recent years, the concept of microsound has taken root in the musical
world. Readily available software generates and manipulates microsound. This
in turn has sparked a great deal of experimentation on the part of musicians.
Judging by the number of compositions that have appeared employing granular
techniques, one could say that these methods have enjoyed a certain vogue. Is a
fascination with the microsonic domain a passing fancy? For some musicians,
this is probably the case. Yet as a resource for expressive sounds, the mine of
microsound is far from exhausted.
Certain timbres produced by granular synthesis have a recognizable signature. These familiar textures can function as memorable units of musical structure. The manipulation of recognizable elements or cliches is an integral part of
composition. Musical material is never entirely new at every level of musical
structure. Even the most advanced realms of musical expression eventually
accumulate a storehouse of cliches, and this accumulation is a natural historical process of music. It is akin to Herbert Brun's concept of ``decay of
information,'' in which familiarity leads to increased legibility at the expense
of freshness (Hamlin and Roads 1985). Although Brun warned that ``understanding, familiarity, and communicativity'' are not necessarily positive aspects, the composer of electronic art music remains in no immediate danger
of being smothered by the hazards of social success.

Conclusion

More than a half century ago, Gabor observed:

Any signal can be expanded in terms of [elementary acoustical quanta] by a process which
includes time analysis and Fourier analysis as extreme cases. (Gabor 1946)

In other words, any sound can be decomposed into constituent particles.

Operations on the Gabor matrix let us transform the particles to create striking
variations of a given sound. Alternatively, one can synthesize entirely new
sounds directly from novel arrangements of elementary particles.
A goal of technology is increased precision and ever-ner control. To make
the Hollerith tabulating machine at the turn of the last century, machinists had
to manufacture parts to the precision of a millimeter. Semiconductor manufacturers now etch microprocessor features down to less than 0.1 of a micron.
The trend toward precision is also reected in our sound tools, which have
passed from wave-oriented to particle-oriented operation, even as the sample
grid has shrunk to just over 5 msec (at a 192 kHz sampling rate).
This book has presented the argument for microsound as a fertile resource in
composition. By now, hundreds of compositions employ granular synthesis and
other particle techniques. The raw material of this music consists of grains and
globules scattered in sonically transparent, diaphanous, or opaque textures.
These textures can be molded, stretched, and combined with others into supple
morphologies. By means of these materials, musical development can take
place not only in the large intervals of pitch and rhythm, but also in microvariations of amplitude, duration, timbre, density, and spatial position.
Perhaps more important than the particles themselves are the sonic brushes
we use to paint them on the canvas of time and frequency. These brushes are
computer programs. By connecting these programs with interactive real-time

350

Chapter 9

controllers, we have built particle synthesis instruments for virtuoso performance, not only for onstage, but also for studio use.
In computer graphics, three-dimensional animation programs incorporate
sophisticated algorithms for scattering particles. These emulate physical models
of ow, perturbation, and collision. In the domain of sound, we can also apply
physical models to regulate the ow of particles. But we should not be limited
to emulations of reality. As stated at the beginning of the book, the computer's
artistic power derives from its ability to model fantasies as well as reality.
Creating sonic fantasies begins with recording, letting us ``photograph'' real
sounds and store their images on tape. The techniques of montagecutting,
splicing, and mixingare essentially manipulations of time structure. As a
pioneer in electronic composition once observed:
If there is one dimension of the music totality, one component which originally led composers to the electronic medium, it was and is the temporal domain. . . . Those who originally turned to electronic tape were obviously attracted to the element of control. After all,
the tape was not a source of sound. Tape is for storage. You can, however, control time as a
measurable distance of tape. Here we are talking about rhythm in every sense of the word.
Not only durational rhythm, but also the time rate of changes of register, of timbre, of
volume, and of those many musical dimensions that were unforeseen until we tried to nd
out how we heard and how we could structure the temporal. (Babbitt 1988)

Discs and semiconductors have largely superseded tape. Mixing and editing
software has replaced the splicing block. A fundamental capability of this software is the zooming in and out on multiple time scales. These tools let us work
at the limits of auditory phenomena, from microsurgery on individual sample
points, to the global rearrangement of sound masses. In the intermediate time
scales, we can edit each sound until it has exactly the right duration, density,
spectral weight, amplitude envelope, and spatial prole. The timing of each
transition can be adjusted precisely.
Much work remains in rethinking existing signal processing paradigms to
take into account the micro time domain. The question is how can researchers
integrate windowing, scattering, and density manipulations with other signal
processing operations on multiple time scales?
Microacoustic phenomena are still not fully understood. Likewise, in science,
the study of granular processes has emerged as a distinct scientic discipline,
with a focus on the interaction of granular streams confronted by external
forces and objects (Behringer and Harmann 1998). At the same time, we see
an emerging science of disordered systems, phase transitions, intermittencies,

351

Conclusion

and particle simulations. These too may serve as fertile analogies for musical
processes.
Guidebooks to the sonic territories, including the present one, are incomplete. Few maps exist, and shortcuts are scarce. So we base our composition
strategy on a heuristic search for remarkable sound objects, mesostructures,
and transformations. It is the job of each composer to discover the interrelations between operations on the micro time scale and their perceptual eects
at other time scales.
Listeners are increasingly open to the beauties found in complex transient
elements, particularly when they appear in lush combination textures. The
acceptance of microsounds also reects an increasing sophistication in their
deployment. When a synthesis technique is rst invented, it is not evident how
to best compose with itevery new instrument requires practice. Much experience has been gained with microsound, and every composer who touches this
resource brings new insights to the puzzle.
Gradually this strange terrain will become familiar as the storehouse of signature gestures accumulates. The understanding shared by only a small circle of
composers today will grow more widespread. We need not concern ourselves as
to whether electronic music will evolve into a language as formal as commonpractice harmony (which is not totally formalized), rather, it is our destiny to
enjoy our newfound land, to invent materials and codes, and to revel in creative
freedom.

This page intentionally left blank

References

Allen, J. B., and L. R. Rabiner. 1977. ``A unied approach to short-time Fourier analysis and synthesis.'' Proceedings of the IEEE 65: 15581564.
American Technology Corporation. 1998. ``HyperSonic sound.'' Internet: www.atcsd.
com.
Apel, K. 1972. Harvard Dictionary of Music. Cambridge, Massachusetts: Harvard University Press.
Arb, D. 1990. ``In the intimacy of a sound.'' In S. Arnold and G. Hair, eds. Proceedings
of the 1990 International Computer Music Conference. pp. 4345.
Arb, D. 1991. ``Analysis, transformation, and resynthesis of musical sounds with the
help of a time-frequency representation.'' In G. De Poli, A. Piccialli, and C. Roads,
eds. 1991. Representations of Musical Signals. Cambridge, Massachusetts: MIT Press.
pp. 87118.
Arb, D. 1998. ``Dierent ways to write digital audio eects programs.'' In B. Garau
and R. Loureiro, eds. Proceedings 98 Digital Audio Eects Workshop. Barcelona: Pompeu Fabra University. pp. 188191.
Arb, D., and N. Delprat. 1992. ``Sound Mutations, a program to transform musical
sounds.'' In A. Strange, ed. Proceedings of the 1992 International Computer Music Conference. San Francisco: International Computer Music Association. pp. 4423.
Arb, D., and N. Delprat. 1993. ``Musical transformations through modications of
time-frequency images.'' Computer Music Journal 17(2): 6672.
Arb, D., and N. Delprat. 1998. ``Selective transformation using a time-frequency representation: an application to the vibrato modication.'' Preprint 4652 (P5S2). Presented at the 104th Convention. New York: Audio Engineering Society.
Arb, D., J. Dudon, and P. Sanchez. 1996. ``WaveLoom, logiciel d'aide a la creation
de disques photosoniques.'' Internet: www.ircam.fr/equipes/repmus/jim96/actes/arb/
waveloom.html.

354

References
Arkani-Hamed, N., S. Dimopoulos, and G. Dvali. 2000. ``The universe's unseen
dimensions.'' Scientic American 283(2): 6269.
Arneodo, A., F. Argoul, E., J. Elezgaray, and J.-F. Muzy. 1995. Ondelettes, multifractales, et turbulences. Paris: Diderot Editeur.
Babbitt, M. 1962. ``Twelve-tone rhythmic structure and the electronic medium.'' Perspectives of New Music 1(1). Reprinted in B. Boretz and E. Cone, eds. 1972. Perspectives
on Centemporary Music Theory. New York: W. W. Norton. pp. 148179.
Babbitt, M. 1988. ``Composition in the electronic medium.'' In F. Roehmann and
F. Wilson, eds. 1988. The Biology of Music-making. Saint Louis: MMB Music. pp. 208
213.
ber die Bedeutung der Ausgleichsvorgange in der Akustik.''
Backhaus, J. 1932. ``U
Zeitschrift fur technische Physik 13(1): 3146.
Backus, J. 1962. ``die Reihe: a scientic evaluation.'' Perspectives of New Music 1(1):
160.
Backus, J. 1969. The Acoustical Foundations of Music. New York: Norton.
Bacry, A., A. Grossman, and J. Zak. 1975. ``Proof of the completeness of the lattice
states in the kq representation.'' Physical Review B12: 1118.
Barlow, C. 1997. ``On the spectral analysis of speech for subsequent resynthesis by
acoustic instruments.'' In F. Barriere and G. Bennett, eds. Analyse en Musique Electroacoustique. Bourges: Editions Mnemosyne. pp. 276283.
Barnwell, T., and C. Richardson. 1995. ``The discrete-time wavelet transformation and
audio coding.'' Preprint 4047 (A-2). Presented at the 99th Convention of the Audio
Engineering Society. New York: Audio Engineering Society.
Barrett, N. 1997. ``Structuring processes in electroacoustic music.'' Ph.D. thesis. London: City University.
Barrett, N. 1998. ``Little Animalscompositional ideas, a brief summary.'' Unpublished
manuscript.
Barrett, N. 1999. ``Little Animals: compositional structuring processes.'' Computer Music Journal 23(2): 1118.
Bartetzki, A. 1997a. ``CMask: ein stochastischer Eventgenerator fur Csound.'' Mitteilunger 26. Berlin: DegeM.
Bartetzki, A. 1997b. ``Csound score generation and granular synthesis with Cmask.''
Internet: www.kgw-tu-berlin.de/@abart/CMaskPaper/cmask-article.html.
Barzun, J. 1961. ``The request for the loan of your ears.'' In H. Russcol. 1972. The Liberation of Sound. Englewood Clis: Prentice-Hall. pp. ixxii.

355

References
Bar-Joseph, Z., D. Lischinski, M. Werman, S. Dubnov, and R. El-Yaniv. 1999.
``Granular synthesis of sound textures using statistical learning.'' In J. Fung, ed. Proceedings of the 1999 International Computer Music Conference. pp. 178181.
Bass, S., and T. Goeddel. 1981. ``The ecient digital implementation of subtractive
music synthesis.'' IEEE Micro 1(3): 2437.
Bastiaans, M. 1980. ``Gabor's expansion of a signal into Gaussian elementary signals.''
Proceedings of the IEEE 68: 538539.
Bastiaans, M. 1985. ``On the sliding-window representation of signals.'' IEEE Transactions on Acoustics, Speech, and Signal Processing ASSP-33(4): 868873.
Bastiaans, M., and M. Geilen. 1996. ``On the discrete Gabor transform and the discrete
Zak transform.'' Signal Processing 49(3): 151166.
Bayle, F. 1989. ``La musique acousmatique ou l'art des sons projetes.'' Paris: Encyclopedia Universalis.
Bayle, F. 1993. Musique Acousmatique. Paris: Institut National de l'Audiovisuel/Groupe
de Recherches Musicales et Buchet/Chastel.
Beauchamp, J. 1969. ``A computer system for time-variant harmonic analysis and synthesis of musical tones.'' In H. von Foerster and J. Beauchamp, eds. Music by Computers. New York: Wiley. pp. 1962.
Beauchamp, J. 1975. ``Analysis and synthesis of cornet tones using nonlinear interharmonic relationships.'' Journal of the Audio Engineering Society 23(10): 718795.
Beekman, I. 16041634. Journal tenu par Isaac Beekman de 1604 a 1634. Four volumes.
C. de Waard, ed. 1953. The Hague.
Behles, G., S. Starke, and A. Roble. 1998. ``Quasi-synchronous and pitch-synchronous
granular sound processing with Stampede II.'' Computer Music Journal 22(2): 4451.
Behringer, R., and H. J. Herrmann, eds. 1998. Granular Matter. Volume 1. Berlin:
Springer-Verlag. ( journal).
Benade, A. 1990. Fundamentals of Musical Acoustics. New York: Dover Publications.
Originally published 1976.
Bencina, R. 2000. Audiomulch software. Internet: www.audiomulch.com.
Bennett, G., and X. Rodet. 1989. ``Synthesis of the singing voice.'' In M. Mathews and
J. Pierce, eds. Current Directions in Computer Music Research. Cambridge, Massachusetts: MIT Press. pp. 1944.
Berg, P. 1978. ``A user's manual for SSP.'' Utrecht: Institute of Sonology.
Berkley, S. 1994. ``QuickMQ: a software tool for the modication of time-varying spectrum analysis les.'' M.S. thesis. Hanover: Department of Music, Dartmouth College.

356

References
Blauert, J. 1997. Spatial Hearing. Cambridge, Massachusetts: MIT Press.
Bode, H. 1984. ``History of electronic sound modication.'' Journal of the Audio Engineering Society 32(10): 730739.
Boerger, G. 1965. ``Die Lokalisation von Gausstonen.'' Doctoral dissertation. Berlin:
Technische Universitat, Berlin.
Boulanger, R. 2000. ed. The Csound Book. Cambridge, Massachusetts: MIT Press.
Boulez, P. 1960. ``Form.'' Darmstadt lecture, reprinted in P. Boulez, 1986. Orientations.
London: Faber and Faber. pp. 9096. Translated by M. Cooper.
Bowcott, P. 1989. ``Cellular automation as a means of high level compositional control
of granular synthesis.'' In T. Wells and D. Butler, eds. 1989. Proceedings of the 1989
International Computer Music Conference. San Francisco: Computer Music Association.
pp. 5557.
Boyer, F., and R. Kronland-Martinet. 1989. ``Granular resynthesis and transformation
of sounds through wavelet transform analysis.'' In T. Wells and T. Butler, eds. Proceedings of the 1989 International Computer Music Conference. San Francisco: International Computer Music Association. pp. 5154.
Brandenburg, K., and M. Bosi. 1997. ``Overview of MPEG audio: current and future
standards for low-bit-rate audio coding.'' Journal of the Audio Engineering Society
45(1/2): 421.
Bristow-Johnson, R. 1995. ``A detailed analysis of a time-domain formant-corrected
pitch-shifting algorithm.'' Journal of the Audio Engineering Society 43(5): 340352.
Brun, H. 1970. ``From musical ideas to computers and back.'' In H. Lincoln, ed. The
Computer and Music. Ithaca: Cornell University Press. pp. 2341.
Brun, H. 1983. Compositions. Program notes with long-play recording Nonsequitur 13.
Champaign: Non Sequitur Records.
Budon, O. 2000. ``Composing with objects, networks, and time scales: an interview with
Horacio Vaggione.'' Computer Music Journal 24(3): 922.
Bumgardner, J. 1997. Syd 1.0 User Manual. Internet: www.jbum.com/jbum.
Buser, P., and M. Imbert. 1992. Audition. Cambridge, Massachusetts: MIT Press.
Butler, D. 1992. The Musician's Guide to Perception and Cognition. New York: Schirmer
Books.
Cage, J. 1937. ``The future of music: credo.'' Lecture reprinted in Cage (1973). pp. 36.
Cage, J. 1959. ``History of experimental music in the United States.'' Lecture reprinted
in Cage (1973). pp. 6775.

357

References
Cage, J. 1973. Silence. Middletown: Wesleyan University Press.
Cahill, T. 1897. U. S. Patent 580, 035.
Cahill, T. 1914. U. S. Patent 1, 107, 261.
Cahill, T. 1917. U. S. Patent 1, 213, 803 and 1, 213, 804.
Cahill, T. 1919. U. S. Patent 1, 295, 691.
Calvet, D., C. Vallee, R. Kronland, and T. Voinier. 2000. ``Descriptif technique du
Cosmophone a 24 voies.'' Internet: cosmophone.in2p3.fr/docs/cosmo24.pdf.
Castonguay, C. 1972. Meaning and Existence in Mathematics. New York: Springer-Verlag.
Castonguay, C. 1973. ``Mathematics and ontology.'' In M. Bunge, ed. The Methodological Unity of Science. Dordrecht: D. Reidel. pp. 1522.
Cavaliere, S., and A. Piccialli. 1997. ``Granular synthesis of musical signals.'' In
C. Roads, S. Pope, A. Piccialli, and G. De Poli, eds. Musical Signal Processing. Lisse:
Swets & Zeitlinger.
Cavaliere, S., G. Evangelista, and A. Piccialli. 1988. ``Synthesis by phase modulation
and its implementation in hardware.'' Computer Music Journal 12(1): 2942.
Cavendish, Margaret Lucas. 1653. Poems and Fancies. London.
Chadabe, J. 1997. Electric Sound. Upper Saddle River: Prentice-Hall.
Chaitin, G. 1998. The Limits of Mathematics. Singapore: Springer-Verlag Singapore.
Chavez, C. 1936. Toward a New Music. Reprinted 1975. New York: Da Capo Press.
Cheng, C. 1996. ``Wavelet signal processing of digital audio with applications in electroacoustic music.'' M.A. thesis. Hanover: Department of Music, Dartmouth College.
Cheng, C. 1997. ``High-frequency compensation of low sample-rate audio les: a wavelet-based spectral excitation algorithm.'' In T. Rikakis, ed. Proceedings of the 1997 International Computer Music Conference. San Francisco: International Computer Music
Association. pp. 458461.
Chion, M. 1982. La musique electroacoustique. Paris: Presses Universitaires de France.
Chou Wen-Chung. 1966. ``A Varese chronology.'' Perspectives of New Music. Reprinted
in B. Boretz and E. Cone, eds. 1971. Perspectives on American Composers. New York:
Norton. pp. 5558.
Chowning, J. 1973. ``The synthesis of complex audio spectra by means of frequency
modulation.'' Journal of the Audio Engineering Society 21(7): 526534. Reprinted in
C. Roads and J. Strawn, eds. 1985. Foundations of Computer Music. Cambridge,
Massachusetts: MIT Press. pp. 629.

358

References
Christensen, E. 1996. The Musical Timespace. Aalborg: Aalborg University Press.
Clark, M. 1993. ``Audio technology in the United States to 1943 and its relationship to
magnetic recording.'' Preprint 3481(H21). New York: Audio Engineering Society.
Clarke, M. 1996. ``Composing at the intersection of time and frequency.'' Organised
Sound 1(2): 107117.
Clarke, M. 1998. ``Extending Contacts: the concept of unity in computer music.'' Perspectives of New Music 36(1): 221239.
Clozier, C. 1998. ``Composition-diusion/interpretation en musique electroacoustique.''
In F. Barriere and G. Bennett, eds. Composition/Diusion en Musique Electroacoustique.
Bourges: Editions Mnemosyne. pp. 52101.
Cochran, W. 1973. The Dynamics of Atoms in Crystals. London: Edward Arnold.
Coelho, V., ed. 1992. Music and Science in the Age of Galileo. Dordrecht: Kluwer Academic Publishers.
Cogan, R. 1984. New Images of Musical Sound. Cambridge, MA: Harvard University
Press.
Cohen, H. 1984. Quantifying Music. Dordrecht: D. Reidel Publishing Company.
Cook, P. 1996. ``Physically informed sonic modeling (PhISM): percussive synthesis.'' In
L. Ayers and A. Horner, eds. Proceedings of the 1996 International Computer Music
Conference. pp. 228231.
Cook, P. 1997. ``Physically informed sonic modeling (PhISM): synthesis of percussive
sounds.'' Computer Music Journal 21(3): 3849.
Cooley, J., and J. Tukey. 1965. ``An algorithm for the machine computation of complex
Fourier series.'' Mathematical Computation 19: 297301.
Cope, D. 1996. Experiments in Musical Intelligence. Madison: A-R Editions.
Correa, J., E. Miranda, and J. Wright. 2000. ``Categorising complex dynamic sounds.''
Internet: www.nyrsound.com.
Cowell, H. 1930. New Musical Resources. New York: A. A. Knopf. Reprinted 1996.
Cambridge, England: Cambridge University Press.
Craven, P. G., and M. A. Gerzon. 1996. ``Lossless coding for audio discs.'' Journal of
the Audio Engineering Society 44(9): 706720.
Crawford, F. 1968. Waves. Berkeley Physics Course Volume 3. New York: McGraw-Hill.
Dahlhaus, C. 1970. ``Aesthetische Probleme der elektronischen Musik.'' In F. Winckel,
ed. Experimentelle Musik. Berlin: Mann-Verlag.

359

References
D'Allessandro, C., and X. Rodet. 1989. ``Synthese et analyse-synthese par fonctions
d'ondes formantiques.'' Journal Acoustique 2: 163169.
Davies, H. 1964. ``Die Reihe Reconsidered1.'' Composer 14: 20.
Davies, H. 1965. ``Die Reihe Reconsidered2.'' Composer 16: 17.
de Broglie, L. 1945. Ondes corpuscules mecanique ondulatoire. Paris: Albin Michel.
de Campo, A. 1998. ``Using recursive diminution for the synthesis of micro time events.''
Unpublished manuscript.
de Campo, A. 1999. SuperCollider Tutorial Workshop. Documentation provided with
the program. Austin: AudioSynth.
Delprat, N., P. Guillemain, and R. Kronland-Martinet. 1990. ``Parameter estimation for
non-linear resynthesis methods with the help of a time-frequency analysis of natural
sounds.'' In S. Arnold and G. Hair, eds. Proceedings of the 1990 International Computer
Music Conference. pp. 8890.
De Poli, G., and A. Piccialli. 1988. ``Forme d'onda per la sintesi granulare sincronica.''
In D. Tommassini, ed. 1988. Atti di VII Colloquio di Informatica Musicale. Rome:
Associazione Musica Verticale. pp. 7075.
De Poli, G., and A. Piccialli. 1991. ``Pitch-synchronous granular synthesis.'' In G. De
Poli, A. Piccialli, and C. Roads, eds. Representations of Musical Signals. Cambridge,
Massachusetts: MIT Press. pp. 187219.
de Reydellet, J. 1999. Personal communication.
Desantos, S. 1997. ``Acousmatic morphology: an interview with Francois Bayle.'' Computer Music Journal 21(3): 1119.
Di Scipio, A. 1990. ``Composition by exploration of nonlinear dynamical systems.'' In
S. Arnold and G. Hair, eds. Proceedings of the 1990 International Computer Music
Conference. San Francisco: International Computer Music Association. pp. 324327.
Di Scipio, A. 1994. ``Formal processes of timbre composition challenging the dualistic
paradigm of computer music.'' Proceedings of ISEA, from Internet: www.uiah./
bookshop/isea_proc/nextgen/j/19.html.
Di Scipio, A. 1995. ``Da Concret PH a GENDY301, modelli compositivi nella musica
elletroacustica de Xenakis.'' Sonusmateriali per la musica contemporanea 7(13): 61
92.
Di Scipio, A. 1997a. ``The problem of 2nd-order sonorities in Xenakis' electroacoustic
music.'' Organised Sound 2(3): 165178.
Di Scipio, A. 1997b. ``Interactive micro-time sound design.'' Journal of Electroacoustic
Music 10: 48.

360

References
Di Scipio, A. 1998. ``Scienza e musica dei quanti acustici: l'eredita di Gabor.'' Il Monocordo 3(6): 6178.
Dodge, C., and T. Jerse. 1997. Computer Music: Synthesis, Composition, and Performance. Second edition. New York: Schirmer.
Dolson, M. 1983. ``A tracking phase vocoder and its use in the analysis of ensemble
sounds.'' Ph.D. dissertation. Pasadena: California Institute of Technology.
Dolson, M. 1985. ``Recent advances in musique concrete at CARL.'' In B. Truax, ed.
Proceedings of the 1985 International Computer Music Conference. San Francisco:
International Computer Music Association. pp. 5560.
Dolson, M. 1986. ``The phase vocoder: a tutorial.'' Computer Music Journal 10(4): 14
27.
Dolson, M., and R. Boulanger. 1985. ``New directions in the musical use of resonators.''
Unpublished manuscript.
Doughty, J., and W. Garner. 1947. ``Pitch characteristics of short tones I: two kinds of
pitch thresholds.'' Journal of Experimental Psychology 37: 351365.
Douglas, A. 1957. The Electrical Production of Music. New York: Philosophical Library.
Dudon, J., and D. Arb. 1990. ``Synthese photosonique.'' Actes du 1er congres
d'Acoustique, colloque C2. Marseille. pp. 845848
Dutilleux, P., A. Grossmann, and R. Kronland-Martinet. 1988. ``Application of the
wavelet transform to the analysis, transformation, and synthesis of musical sounds.''
Preprint 2727 (A2). Presented at the 85th Convention. New York: Audio Engineering
Society.
Eckel, G. 1990. ``A signal editor for the IRCAM Musical Workstation.'' In S. Arnold
and G. Hair, eds. Proceedings of the 1990 International Computer Music Conference. San
Francisco: International Computer Music Association. pp. 6971.
Eckel, G., M. Rocha-Iturbide, and B. Becker. 1995. ``The development of GiST, a
granular synthesis toolkit based on an extension of the FOF Generator.'' In R. Bidlack,
ed. Proceedings of the 1995 International Computer Music Conference. San Francisco:
International Computer Music Association. pp. 296302.
Eimert, H. 1955. ``What is electronic music.'' die Reihe 1:110. English edition 1958.
Bryn Mawr: Theodore Presser Company.
Einstein, A. 1952. Relativity: The Special and the General Theory. Fifteenth edition of
the original published in 1916. New York: Three Rivers Press.
Elmore, W., and M. Heald. 1969. Physics of Waves. New York: Dover.

361

References
Endrich, A., Coordinator. 2000. Composer's Desktop Project. 12 Goodwood Way,
Cepen Park South, Chippenham Wiltshire SN14 0SY, United Kingdom. archer@
trans4um.demon.co.uk.
Erbe, T. 1995. SoundHack User's Manual. Oakland: Mills College.
Erne, M. 1998. ``Embedded audio compression using wavelets and improved psychoacoustic models.'' In B. Garau and R. Loureiro, eds. Proceedings 98 Digital Audio
Eects Workshop. Barcelona: Pompeu Fabra University. pp. 147150.
Evangelista, G. 1991. ``Wavelet transforms that we can play.'' In G. De Poli, A. Piccialli, and C. Roads, eds. 1991. Representations of Musical Signals. Cambridge, Massachusetts: MIT Press. pp. 119136.
Evangelista, G. 1997. ``Wavelet representation of musical signals.'' In Roads, et al. eds.
Musical Signal Processing. Lisse: Swets and Zeitlinger. pp. 126153.
Evangelista, G., and S. Cavaliere. 1998. ``Dispersive and pitch-synchronous processing
of sounds.'' In B. Garau and R. Loureiro, eds. Proceedings 98 Digital Audio Eects
Workshop. Barcelona: Pompeu Fabra University. pp. 232236.
Fairbanks, G., W. Everitt, and R. Jaeger. 1954. ``Method for time or frequency compression-expansion of speech.'' Institute of Radio Engineers Transactions on Audio AV
2(1): 712.
Feichtinger, H., and T. Strohmer, eds. 1998. Gabor Analysis and Algorithms. Boston:
Birkhauser.
Fields, K. 1998. Personal communication.
Fitz, K., L. Haken, and B. Holloway. 1995. Lemur Pro 4.0.1 documentation. Urbana:
Engineering Research Laboratory, University of Illinois.
Flanagan, J. L., and R. Golden. 1966. ``Phase vocoder.'' Bell System Technical Journal
45: 14931509.
Flanagan, J. L. 1972. Speech Analysis, Synthesis, and Perception. New York: SpringerVerlag.
Fletcher, N., and T. Rossing. 1991. The Physics of Musical Instruments. New York:
Springer-Verlag.
Fokker, A. 1962. ``Wherefore, and Why?'' die Reihe 8. English edition 1968, 6879.
Fourier, L. 1994. ``Jean-Jacques Perrey and the Ondioline.'' Computer Music Journal
18(4): 1825.
Fraisse, P. 1982. ``Rhythm and tempo.'' In D. Deutsch, ed. The Psychology of Music.
Orlando: Academic.

362

References
Freed, A. 1987. ``MacMix: recording, mixing, and signal processing on a personal computer.'' In J. Strawn, ed. 1987. Music and Digital Technology. New York: Audio Engineering Society. pp. 158162.
Freedman, M. D. 1967. ``Analysis of musical instrument tones.'' Journal of the Acoustical Society of America 41: 793806.
Freeman, W. 1991. ``The physiology of perception.'' Scientic American 264(2): 7885.
Freeman, W. 1995. ``Chaos in the central nervous system.'' In F. Ventriglia, ed. Neural
Modeling and Neural Networks. New York: Pergamon Press. pp. 185216.
Gabor, D. 1946. ``Theory of communication.'' Journal of the Institute of Electrical
Engineers Part III, 93: 429457.
Gabor, D. 1947. ``Acoustical quanta and the theory of hearing.'' Nature 159(4044): 591
594.
Gabor, D. 1952. ``Lectures on communication theory.'' Technical Report 238, Research
Laboratory of Electronics. Cambridge, Massachusetts: Massachusetts Institute of Technology.
Galente, F., and N. Sani. 2000. Musica Espansa. Lucca: Ricordi LIM.
Galilei, Galileo. 1623. ``Parable of Sound.'' Quoted in V. Coelho. ``Musical myth and
Galilean science in Giovanni Serodine's Allegoria della Scienza.'' See V. Coelho (1992).
Gardner, M. 1957. Fads and Fallacies in the Name of Science. New York: Dover.
Gardner, W. 1995. ``Ecient convolution without input-output delay.'' Journal of the
Audio Engineering Society 43(3): 127136.
Gassendi, P. 1658. Syntagma Philosophicum.
George, E., and M. Smith. 1992. ``Analysis-by-synthesis/overlap-add sinusoidal modeling applied to the analysis and synthesis of musical tones.'' Journal of the Audio Engineering Society 40(6): 497516.
Gerrard, G. 1989. ``Music4CA Macintosh version of Music IVBF in C.'' Melbourne:
University of Melbourne.
Gerzon, M. 1973. ``Periphony: with-height sound reproduction.'' Journal of the Audio
Engineering Society 21(1): 210.
Gibson, B., C. Jubien, and B. Roden. 1996. ``Method and apparatus for changing the
timbre and/or pitch of audio signals.'' U. S. Patent 5,567,901.
Gleick, J. Chaos. 1988. London: Cardinal.
Goeyvaerts, K. 1955. ``The sound material of electronic music.'' die Reihe 1: 357.
English edition 1958. Bryn Mawr: Theodore Presser Company.

363

References
Gogins, M. 1991. ``Iterated function systems music.'' Computer Music Journal 15(1): 40
48.
Gogins, M. 1995. ``Gabor synthesis of recurrent iterated function systems.'' In R.
Bidlack, ed. Proceedings of the 1995 International Computer Music Conference. San
Francisco: International Computer Music Association. pp. 349350.
Gordon, J. 1996. ``Psychoacoustics in computer music.'' In C. Roads. 1996. The Computer Music Tutorial. Cambridge, Massachusetts: MIT Press. pp. 10531069.
Green, D. 1971. ``Temporal auditory acuity.'' Psychological Review 78(6): 540551.
Grey, J. 1975. ``An exploration of musical timbre.'' Report STANM2. Stanford
University Department of Music.
Grossman, A., M. Holschneider, R. Kronland-Martinet, and J. Morlet. 1987. ``Detection of abrupt changes in sound signals with the help of wavelet transforms.'' Inverse
Problems: Advances in Electronic and Electronic Physics. Supplement 19. San Diego:
Academic Press. pp. 298306.
Hamlin, P., with C. Roads. 1985. ``Interview with Herbert Brun.'' In C. Roads, ed.
Composers and the Computer. Madison: A-R Editions. pp. 115.
Harley, J. forthcoming. The Music of Iannis Xenakis. London: Harwood Academic
Publishers.
Harris, F. 1978. ``On the use of windows for harmonic analysis with the discrete Fourier
transform.'' Proceedings of the IEEE 66(1): 5183.
Hawking, S., and R. Penrose. 1996. ``The nature of space and time.'' Scientic American
248(7): 6065.
Heisenberg, W. 1958. Physics and Philosophy. New York: Harper.
Helmholtz, H. 1885. On the Sensations of Tone. Translated by A. Ellis. New edition
1954. New York: Dover Publications.
Helmuth, M. 1991. ``Patchmix and StochGran: Two Graphical Interfaces.'' In B.
Alphonce and B. Pennycook, eds. Proceedings of the 1991 International Computer Music
Conference. San Francisco: International Computer Music Association. pp. 563565.
Helmuth, M. 1993. ``Granular Synthesis with Cmix and Max.'' In S. Ohteru, ed. Proceedings of the 1993 International Computer Music Conference. San Francisco: International Computer Music Association. pp. 449452.
Helstrom, C. 1966. ``An expansion of a signal in Gaussian elementary signals.'' IEEE
Transactions on Information Theory IT12: 8182.
Hiller, L., and L. Isaacson. 1959. Experimental Music. New York: McGraw-Hill.

364

References
Hiller, L. 1970. ``Music composed with computersa historical survey.'' In H. Lincoln,
ed. The Computer and Music. Ithaca: Cornell University Press. pp. 4296.
Homann, P. 1994. Amalgam aus Kunst und Wissenschaft. Naturwissenschaftliches
Denken im Werk von Iannis Xenakis. Frankfurt: Peter Lang.
Homann, P. 1996. ``Amalgamation of art and science: scientic thought in the work of
Iannis Xenakis.'' Abbreviated version of Homann (1994) in English. Unpublished
manuscript.
Homann, P. 1997. Personal communication.
Holden, A. 1986. Chaos. Princeton: Princeton University Press.
Howard, E. 1996. Personal communication.
Howe, H. S., Jr. 1975. Electronic Music Synthesis. New York: Norton.
Hugo, V. 1862. Les Miserables. Reprinted 1999. Paris: Gallimard.
Ives, C. 1962. in H. Boatwright, ed. Essays Before a Sonata and Other Writings. New
York: W. W. Norton. p. 98.
Jenny, G. 195556. ``Initiation a la lutherie electronique.'' Toute la Radio (September
1955): 28994, (November 1955): 397404, (December 1955): 4559, January 1956):
236, (February 1956) 6772.
Jenny, G. 1958. ``L'Ondioline: conception et realisation.'' Toute la radio.
Jones, D., and T. Parks. 1986. ``Time scale modication of signals using a synchronous
Gabor technique.'' Proceedings of the 1986 ASSP Workshop on Applications of Signal
Processing to Audio and Acoustics. New York: IEEE.
Jones, D., and T. Parks. 1988. ``Generation and combination of grains for music synthesis.'' Computer Music Journal 12(2): 2734.
Kaegi, W. 1967. Was Ist Elektronische Musik. Zurich: Orell Fussli Verlag.
Kaegi, W. 1973. ``A minimum description of the linguistic sign repertoire (part 1).''
Interface 2: 141156.
Kaegi, W. 1974. ``A minimum description of the linguistic sign repertoire (part 2).''
Interface 3: 137158.
Kaegi, W., and S. Tempelaars. 1978. ``VOSIMa new sound synthesis system.'' Journal
of the Audio Engineering Society 26(6): 418426.
Kaiser, G. 1994. A Friendly Guide to Wavelets. Boston: Birkhauser.
Kaku, M. 1995. Hyperspace. New York: Anchor Books.

365

References
Kaler, J. 1997. Cosmic Clouds. New York: Scientic American Library.
Kandinsky, W. 1926. Point et ligne sur plan. 1991 edition, Paris: Gallimard.
Kargon, R. H. 1966. Atomism in England from Hariot to Newton. Oxford: Clarendon Press.
Keller, D. 1999. Touch 'n Go. Enhanced compact disc with synthesis data. ES 99002.
New Westminster, British Columbia: Earsay.
Keller, D., and C. Rolfe. 1998. ``The corner eect.'' Proceedings of the XII Colloquium
on Musical Informatics. Gorizia: Centro Polifunzionale di Gorizia. Also see Internet:
www.thirdmonk.com.
Keller, D., and B. Truax. 1998. ``Ecologically-based granular synthesis.'' In M. Simoni,
ed. Proceedings of the 1998 International Computer Music Conference. San Francisco:
International Computer Music Association. pp. 117120.
Kientzle, T. 1998. A Programmer's Guide to Sound. Reading, Massachusetts: AddisonWesley.
Koenig, G. M. 1959. ``Studium im Studio.'' die Reihe 5. English edition 1961. Bryn
Mawr: Theodore Presser Company. pp. 3039.
Koenig, G. M. 1962. ``Commentary on Stockhausen's . . . how time passes . . . on Fokker's wherefore, and why, and on present musical practice as seen by the author.'' die
Reihe 8. English edition 1968. Bryn Mawr: Theodore Presser Company. pp. 8098.
Koenig, R. 1899. Articles in Annalen der Physik 69: 626660, 721738. Cited in Miller
1916, 1935.
Koenigsberg, C. 1991. ``Stockhausen's new morphology of musical time.'' Posted June
1996. Internet: www2.uchicago.edu/ns-acs/ckk/smmt.
Krenek, E. 1955. ``A glance over the shoulders of the young.'' die Reihe 1: 1416. English edition 1958. Bryn Mawr: Theodore Presser Company.
Kronland-Martinet, R. 1988. ``The wavelet transform for the analysis, synthesis, and
processing of speech and music sounds.'' Computer Music Journal 12(4): 1120.
Kronland-Martinet, R., and A. Grossman. 1991. ``Application of time-frequency and
time-scale methods (wavelet transforms) to the analysis, synthesis, and transformation of
sounds.'' In G. De Poli, A. Piccialli, and C. Roads, eds. 1991. Representations of Musical
Signals. Cambridge, Massachusetts: MIT Press. pp. 4585.
Kronland-Martinet, R., Ph. Guilliman, and S. Ystad. 1997. ``Modeling of natural
sounds by time-frequency and wavelet representations.'' Organised Sound 2(3): 179191.
Kronland-Martinet, R., J. Morlet, and A. Grossmann. 1987. ``Analysis of sound patterns through wavelet transforms.'' International Journal on Pattern Recognition and
Articial Intelligence 1(2): 273302.

366

References
Kunt, M. 1981. Traitement numerique des signaux. Paris: Dunod.
Kupper, L. 2000. ``Le temps audio-numerique.'' In C. Clozier and F. Barriere, eds.
Les actes d'academie de musique electroacoustique. Bourges: Institute International de
Musique Electoacoustique de Bourges. pp. 94115.
Kussmaul, C. 1991. ``Applications of wavelets in music: the wavelet function library.''
M. A. thesis. Hanover: Dartmouth College.
Langmead, C. 1995. Perceptual Analysis Synthesis Tool 1.6 Manual. Hanover, New
Hampshire: Dartmouth College.
Learned, R., and A. Willsky. 1993. ``Wavelet packet approach to transient signal classication.'' Technical Report LIDSP2199, Laboratory for Information and Decision
Systems. Cambridge: MIT
Lederman, L., and D. Schramm. 1995. Second edition. From Quarks to the Cosmos. San
Francisco: W. H. Freeman.
Lee, F. 1972. ``Time compression and expansion of speech by the sampling method.''
Journal of the Audio Engineering Society 20(9): 738742.
Lee, A. 1995. ``Csound granular synthesis unit generator.'' In Proceedings of the 1995
International Computer Music Conference. San Francisco: International Computer Music Association. pp. 2301.
Leichtentritt, H. 1951. Musical Form. Englewood Clis: Prentice-Hall.
Lent, K. 1989. ``An ecient method for pitch shifting digitally sampled sounds.'' Computer Music Journal 13(4): 6571.
Lesbros, V. 1995. ``Atelier incremental pour la musique experimentale.'' These doctorat
en Intelligence Articielle. Paris: Universite Paris 8.
Lesbros, V. 1996. ``From images to sounds: a dual representation.'' Computer Music
Journal 20(3): 5969.
Liandrat, J., and F. Moret-Bailly. 1990. ``The wavelet transform: some applications to
uid dynamics and turbulence.'' European Journal of Mech., B/Fluids 9: 119.
Ligeti, G. 1971. ``Fragan und Antworden von mir selbst.'' Melos 12: 509516.
Lippe, C. 1993. ``A musical application of real-time granular sampling using the
IRCAM signal processing workstation.'' In S. Ohteru, ed. Proceedings of the 1993 International Computer Music Conference. San Francisco: International Computer Music
Association. pp. 190193.
Lippman, E. 1992. A History of Western Music Aesthetics. Lincoln: University of
Nebraska Press.

367

References
Lopez, S., F. Mart, and E. Resina. 1998. ``Vocem: an application for real-time granular
synthesis.'' In B. Garau and R. Loureiro, eds. Proceedings 98 Digital Audio Eects
Workshop. Barcelona: Pompeu Fabra University. pp. 219222.
Lorrain, D. 1980. ``A panoply of stochastic `cannons'.'' Computer Music Journal 4(1):
5381. Reprinted in C. Roads. 1989. The Music Machine. Cambridge, Massachusetts:
MIT Press. pp. 351379.
Luce, D. 1963. ``Physical correlates of nonpercussive instrument tones.'' Sc.D. dissertation. Cambridge, Massachusetts: MIT Department of Physics.
Lucretius. 55. De Rerum Natura. In J. Gaskin, ed. 1995. The Epicurean Philosophers.
London: Everyman. pp. 79304.
Maconie, R. 1989. Stockhausen on Music. London: Marion Boyars.
Maher, R., and J. Beauchamp. 1990. ``An investigation of vocal vibrato for synthesis.''
Applied Acoustics 30: 219245.
Malah, D. 1979. ``Time-domain algorithms for harmonic bandwidth reduction and time
scaling of speech signals.'' IEEE Transactions on Acoustics, Speech, and Signal Processing 27(4): 121133.
Mallat, G. 1988. ``Review of multifrequency channel decompositions of images and
wavelet models.'' Technical Report No. 412. New York: New York University, Department of Computer Science.
Mallat, G. 1989. ``A theory for multiresolution signal decomposition: the wavelet representation.'' IEEE Transactions on Pattern Analysis and Machine Intelligence 11(7):
674693.
Mallat, S. G. 1998. A Wavelet Tour of Signal Processing. San Diego: Academic Press.
Marino, G., J.-M. Raczinski, and M.-H. Serra. 1990. ``The new UPIC system.'' In
S. Arnold and G. Hair, eds. Proceedings of the 1990 International Computer Music
Conference. San Francisco: International Computer Music Association. pp. 249252.
Marple, S. 1987. Digital Spectral Analysis. Englewood Clis: Prentice-Hall.
Masri, P., A. Bateman, and N. Canagarajah. 1997a. ``A review of time-frequency representations, with an application to sound/music analysis-resynthesis.'' Organised Sound
2(3): 193205.
Masri, P., A. Bateman, and N. Canagarajah. 1997b. ``The importance of the time-frequency representation for sound/music analysis-resynthesis.'' Organised Sound 2(3):
207214.
Mathews, M. 1969. The Technology of Computer Music. Cambridge, Massachusetts:
MIT Press.

368

References
Mathews, M., and J. Miller. 1965. Music IV Programmer's Manual. Murray Hill: Bell
Telephone Laboratories.
Mathews, M., J. Miller, and E. David, Jr. 1961. ``Pitch synchronous analysis of voiced
sounds.'' Journal of the Acoustical Society of America 33: 179186.
MathWorks, The. 1995. Matlab Reference Guide. Natick: The MathWorks.
McAdams, S. 1982. ``Spectral fusion and the creation of auditory images.'' In M.
Clynes, ed. Music, Mind, and Brain. New York: Plenum.
McAdams, S., and A. Bregman. 1979. ``Hearing musical streams.'' Computer Music
Journal 3(4): 2644. Reprinted in C. Roads and J. Strawn, eds. 1985. Foundations of
Computer Music. Cambridge, Massachusetts: MIT Press. pp. 658698.
McAulay, R., and T. Quatieri. 1986. ``Speech analysis/synthesis based on a sinusoidal
representation.'' IEEE Transactions on Acoustics, Speech, and Signal Processing ASSP
34: 744754.
McCartney, J. 1990. Synth-O-Matic version 0.06 software.
McCartney, J. 1994. Synth-O-Matic version 0.45 software.
McCartney, J., 1996. SuperCollider, A Real-time Sound Synthesis Programming Language. Austin: AudioSynth.
McCartney, J. 1998. ``Continued evolution of the SuperCollider real time synthesis
environment.'' In M. Simoni, ed. Proceedings of the 1998 International Computer
Music Conference. San Francisco: International Computer Music Association. pp. 133
136.
Meijer, P. 1998. ``Auditory wavelets?'' Internet: ourworld.compuserve.com/homepages/
Peter_Meijer/wavelet.htm.
Meridian. 1998. ``Meridian lossless packing: provisional data June 1998.'' Huntington:
Boothroyd-Stuart Meridian.
Mersenne, M. 1636. Harmonie Universelle. Reprinted 1957, translated by Roger E.
Chapman. The Hague: Martinus Nijho.
Meyer, Y. 1994. Les ondelettes: algorithmes et applications. Paris: Armand Colin editeur.
Meyer, Y., S. Jaard, and O. Rioul. 1987. ``L'analyse par ondelettes.'' Pour la Science.
September. pp. 2837.
Meyer-Eppler, W. 1959. Grundlagen und Aufwendungen der Informationstheorie.
Springer-Verlag: Berlin.
Meyer-Eppler, W. 1960. ``Zur Systematik der elektrischen Klangtransformation.''
Darmstadter Beitrage zur Neuen Musik III. Mainz: Schott.

369

References
Miller, D. C. 1916. The Science of Musical Sounds. New York: MacMillan.
Miller, H. 1945. The Air-conditioned Nightmare. New York: New Directions.
Miranda, E. 1998. ``Chaosynth: a cellular automata-based granular synthesiser.'' Internet: website.lineone.net/@edandalex/chaosynt.htm.
Moles, A. 1960. Les musiques experimentales. Zurich: Editions du Cercle de l'Art Contemporain.
Moles, A. 1968. Information Theory and Esthetic Perception. Urbana: University of
Illinois Press.
Moog, R. 1996. ``Build the EM Theremin.'' Electronic Musician 12(2): 8699.
Moon, F. 1987. Chaotic Vibrations. New York: Wiley-Interscience.
Moore, F. R. 1982. ``The computer audio research laboratory at UCSD.'' Computer
Music Journal 6(1): 1829.
Moore, F. R. 1990. Elements of Computer Music. Englewood Clis: Prentice-Hall.
Moorer, J. A. 1977a. ``Editing, mixing, and processing digitized audio waveforms.''
Paper presented at the 57th Convention of the Audio Engineering Society. New York:
Audio Engineering Society.
Moorer, J. A. 1977b. ``Signal processing aspects of computer music.'' Proceeding of the
IEEE 65(8): 11081137. Reprinted in Computer Music Journal 1(1): 437 and in J.
Strawn, ed. 1985. Digital Audio Signal Processing: An Anthology. Madison: A-R
Editions.
Moorer, J. A. 1978. ``The use of the phase vocoder in computer music applications.''
Journal of the Audio Engineering Society 26(1/2): 4245.
Morawska-Bungler, M. 1988. Schwingende Elektronen. Cologne: P. J. Tonger.
Morris, R. 1987. Composition with Pitch-Classes. New Haven: Yale University Press.
Nahin, P. 1996. The Science of Radio. Woodbury: American Institute of Physics.
Navarro, R., A. Tabernero, and G. Cristobal. 1995. ``Image representation with Gabor
wavelets and its applications.'' In P. Hawkes, ed. Advances in Imaging and Electron
Physics. Orlando: Academic Press.
Nelson, G. 1997. ``Wind, Sand, and Sea Voyages: an application of granular synthesis
and chaos to musical composition.'' www.timara.oberlin.edu/people/%7Egnelson/
papers/Gola/gola.pdf.
Nelson, Jon Christopher. 1996. They Wash Their Ambassadors in Citrus and Fennel. On
the compact disc Cultures Electroniques 9. LCD 278060/61. Bourges: Serie Bourges/
Unesco/Cime.

370

References
Neve, R. 1992. ``Rupert Neve of Amek replies.'' Studio Sound 34(3): 2122.
Newland, D. 1994. ``Harmonic and musical wavelets.'' Proceedings of the Royal Society
of London A 444: 605620.
Norris, M. 1997. SoundMagic 1.0.3 Documentation. Wellington, New Zealand: Michael
Norris.
Nuttall, A. 1981. ``Some windows with very good sidelobe behavior.'' IEEE Transactions on Acoustics, Speech, and Signal Processing ASSP29(1): 8491.
Nyquist, H. 1928. ``Certain topics in telegraph transmission theory.'' Transactions of the
American Institute of Electrical Engineers 4.
Olson, H. F. 1957. Acoustical Engineering. New York: D. Van Nostrand. Reprinted
1991. Philadephia: Professional Audio Journals.
Oohashi, T., E. Nishina, N. Kawai, Y. Fuwamoto, and H. Imai. 1991. ``High frequency
sound above the audible range aects brain electric activity and sound perception.''
Preprint 3207(W1). Presented at the 91st Convention of the Audio Engineering Society.
New York: Audio Engineering Society.
Oohashi, T., E. Nishina, Y. Fuwamoto, and N. Kawai. 1993. ``On the mechanism of
hypersonic eect.'' In S. Ohteru, ed. Proceedings of the 1993 International Computer
Music Conference. San Francisco: International Computer Music Association. pp. 432
434.
Orcalli, A. 1993. Fenomenologia della Musica Sperimentale. Potenza: Sonus Edizioni
Musicali.
Orton, R., A. Hunt, and R. Kirk. 1991. ``Graphical control of granular synthesis using
cellular automata and the Freehand program.'' In B. Alphonce and B. Pennycook, eds.
Proceedings of the 1991 International Computer Music Conference. San Francisco:
International Computer Music Association. pp. 416418.
Otis, A., G. Grossman, and J. Cuomo. 1968. ``Four sound-processing programs for the
Illiac II computer and D/A converter.'' Experimental Music Studios Technical Report
Number 14. Urbana: University of Illinois.
Pape, G. 1998. Personal communication.
Piccialli, A., S. Cavaliere, I. Ortosecco, and P. Basile. 1992. ``Modications of natural
sounds using a pitch synchronous technique.'' In A. Piccialli, ed. Proceedings of the International Workshop on Models and Representations of Musical Signals. Napoli: Universita di Napoli Federico II.
Pierce, A. D. 1994. Acoustics: An Introduction to Its Physical Principles and Applications.
Woodbury, New York: Acoustical Society of America.

371

References
Pierce, J. R. 1983. The Science of Musical Sound. New York: W. H. Freeman.
Pines, D. 1963. Elementary Excitations in Solids. New York: W. A. Benjamin.
Polansky, L., and T. Erbe. 1996. ``Spectral mutation in SoundHack.'' Computer Music
Journal 20(1): 92101.
Pompei, F. J. 1998. ``The use of airborne ultrasonics for generating audible sound
beams.'' Preprint 4853 (I5). New York: Audio Engineering Society.
Pompei, F. J. 1999. ``The use of airborne ultrasonics for generating audible sound
beams.'' Journal of the Audio Engineering Society 47(9): 726731.
Pope, S. T. 1997. Sound and Music Processing in SuperCollider. Santa Barbara:
www.create.ucsb.edu.
Portno, M. 1976. ``Implementation of the digital phase vocoder using the fast Fourier
transform.'' IEEE Transactions on Acoustics, Speech and Signal Processing 24(3): 243
248.
Portno, M. 1978. ``Time-scale modication of speech based on short-time Fourier
analysis.'' Sc.D. dissertation. Cambridge, Massachusetts: MIT Department of Electrical
Engineering and Computer Science.
Portno, M. 1980. ``Time-frequency representation of digital signals and systems based
on short-time Fourier analysis.'' IEEE Transactions on Acoustics, Speech, and Signal
Processing 28: 5569.
Portno, M. 1981. ``Time-scale modication of speech based on short-time Fourier
analysis.'' IEEE Transactions on Acoustics, Speech, and Signal Processing 29(3): 374
390.
Pound, E. 1934. Quoted in R. Murray Schafer, editor, 1977, Ezra Pound and Music: the
Complete Criticism. New York: New Directions.
Pranger, M. 1999. MarcoHack Version 1 Manual. Internet: www.koncon.nl/MarcoHack.
Quate, C. 1979. ``The acoustic microscope.'' Scientic American. Reprinted in
1998, Science's Vision: The Mechanics of Sight. New York: Scientic American. pp. 31
39.
Rabiner, L., and B. Gold. 1975. Theory and Application of Digital Signal Processing.
Englewood Clis: Prentice-Hall.
Rabiner, L., J. Cooley, H. Helms, L. Jackson, J. Kaiser, C. Rader, R. Schafer, K. Steiglitz, and C. Weinstein. 1972. ``Terminology in digital signal processing.'' IEEE Transactions on Audio and Electroacoustics AU20: 3227.

372

References
Raczinski, J.-M., and G. Marino. 1988. ``A real time synthesis unit.'' In C. Lischka and
J. Fritsch, eds. Proceedings of the 1988 International Computer Music Conference. San
Francisco: International Computer Music Association. pp. 90100.
Raczinski, J.-M., G. Marino, and M.-H. Serra. 1991. ``New UPIC system demonstration.'' In B. Alphonce and B. Pennycook, eds. Proceedings of the 1991 International
Computer Music Conference. San Francisco: International Computer Music Association.
pp. 567570.
Ramarapu, P., and R. Maher. 1997. ``Methods for reducing audible artefacts in a
wavelet-based broad-band denoising system.'' Journal of the Audio Engineering Society
46(3): 178190.
Read, O., and W. Welch. 1976. From Tin Foil to Stereo: Evolution of the Phonograph.
Indianapolis: Howard Sams.
Reder, L., and J. S. Gordon. 1997. ``Subliminal perception: nothing special, cognitively
speaking.'' In J. Cohen and J. Schooler, eds. Cognitive and Neuropsychological
Approaches to the Study of Consciousness. Hillsdale, New Jersey: Lawrence Erlbaum
Associates. pp. 125134.
Reeves, W. 1983. ``Particle systemsa technique for modeling a class of fuzzy objects.''
ACM Transactions on Graphics 2(2): 359376.
Rhea, T. 1972. ``The evolution of electronic musical instruments in the United States.''
Ph.D. dissertation. Nashville: Peabody College.
Rhea, T. 1984. ``The history of electronic musical instruments.'' In T. Darter, ed. 1984.
The Art of Electronic Music. New York: Quill. pp. 163.
Ricci, A. 1997. SoundMaker software. Distributed by MicroMat Computer Systems.
Risset, J.-C. 1966. ``Computer study of trumpet tones.'' Murray Hill: Bell Telephone
Laboratories.
Risset, J.-C. 1969. An Introductory Catalog of Computer Synthesized Sounds (With
Sound Examples). Murray Hill: Bell Laboratories. Reprinted 1995 with the compact
disc The Historical CD of Digital Sound Synthesis, Computer Music Currents 13, Wergo
20332. Mainz: Wergo Schallplatten.
Risset, J.-C. 1985. ``Computer music experiments: 1964. . .'' Computer Music Journal
9(1): 1118. Reprinted in C. Roads, ed. 1989. The Music Machine. Cambridge, Massachusetts: MIT Press. pp. 6774.
Risset, J.-C. 1989a. ``Paradoxical sounds.'' In M. Mathews and J. Pierce, eds. 1989.
Current Directions in Computer Music Research. Cambridge, Massachusetts: The MIT
Press. pp. 149158.

373

References
Risset, J.-C. 1989b. ``Additive synthesis of inharmonic tones.'' In M. Mathews and
J. Pierce, eds. 1989. Current Directions in Computer Music Research. Cambridge,
Massachusetts: The MIT Press. pp. 159163.
Risset, J.-C. 1991. ``Timbre analysis by synthesis: representations, imitations, and
variants for musical composition.'' In G. De Poli, A. Piccialli, and C. Roads, eds.
1991. Representations of Musical Signals. Cambridge, Massachusetts: MIT Press. pp. 7
43.
Risset, J.-C. 1992. ``Composing sounds with computers.'' In J. Paynter, T. Howell, R.
Orton, and P. Seymour, eds. Companion to Contemporary Musical Thought. London:
Routledge. pp. 583621.
Risset, J.-C. 1996. ``Denition de la musique electroacoustique.'' In G. Bennett, C.
Clozier, S. Hanson, C. Roads, and H. Vaggione, eds. Esthetique et Musique Electroacoustique. Bourges: Mnemosyne. pp. 8284.
Risset, J.-C. 1997. ``Problemes d'analyse : quelque cles pour mes premieres pieces.'' In
F. Barriere and G. Bennett, eds. Analyse en Musique Electroacoustique. Bourges:
Editions Mnemosyne. pp. 169177.
Risset, J.-C. 1998. ``Examples of the musical use of digital audio eects.'' In B. Garau
and R. Loureiro, eds. Proceedings 98 Digital Audio Eects Workshop. Barcelona:
Pompeu Fabra University. pp. 254259.
Risset, J.-C. 1999a. Personal communication.
Risset, J.-C. 1999b. Overview of research at the Laboratoire de Mecanique et d'Acoustique. Internet: omicron.cnrs_mrs.fr.
Risset, J.-C., and D. Wessel. 1982. ``Exploration of timbre by analysis and synthesis.'' In
D. Deutsch, ed. 1982. The Psychology of Music. Orlando: Academic. pp. 2558.
Roads, C. 1975. ``Computer music studies 19741975.'' Unpublished manuscript. 44
pages.
Roads, C. 1976. ``A systems approach to composition.'' Honors thesis. La Jolla: University of California, San Diego.
Roads, C. 1978a. ``Automated granular synthesis of sound.'' Computer Music Journal
2(2): 6162. Revised and updated version printed as ``Granular synthesis of sound'' in C.
Roads and J. Strawn, eds. 1985. Foundations of Computer Music. Cambridge, Massachusetts: MIT Press. pp. 145159.
Roads, C. 1978b. ``Interview with Gottfried Michael Koenig.'' Computer Music Journal
2(3): 1115. Reprinted in C. Roads and J. Strawn, eds. 1985. Foundations of Computer
Music. Cambridge, Massachusetts: MIT Press. pp. 568580.
Roads, C., ed. 1985a. Composers and the Computer. Madison: A-R Editions.

374

References
Roads, C. 1985b. ``The realization of nscor.'' In C. Roads, ed. 1985. Composers and the
Computer. Madison: A-R Editions. pp. 140168.
Roads, C. 1985c. ``Granular synthesis of sound.'' In C. Roads and J. Strawn, eds. 1985.
Foundations of Computer Music. Cambridge, Massachusetts: MIT Press. pp. 145159.
Roads, C. 1985d. ``Grammars as representations for music.'' In C. Roads and J. Strawn,
eds. 1985. Foundations of Computer Music. Cambridge, Massachusetts: MIT Press.
pp. 403442.
Roads, C. 1987. ``Experiences with computer-assisted composition.'' Translated as
``Esperienze di composizione assistata da calculatore.'' In S. Tamburini and M. Bagella,
eds. I Proli del Suono. Rome: Musica Verticale and Galzeramo. pp. 173196.
Roads, C. 1991. ``Asynchronous granular synthesis.'' In G. De Poli, A. Piccialli, and
C. Roads, eds. 1991. Representations of Musical Signals. Cambridge, Massachusetts:
MIT Press. pp. 143185.
Roads, C. 1992a. ``Composition with machines.'' In J. Paynter, T. Howell, R. Orton,
and P. Seymour, eds. Companion to Contemporary Musical Thought. London: Routledge. pp. 399425.
Roads, C. 1992b. ``Musical applications of advanced signal transformations.'' In
A. Piccialli, ed. Proceedings of the Capri Workshop on Models and Representations of
Musical Signals. Naples: University of Naples Federico II, Department of Physics.
Roads, C. 1992c. ``Synthulate.'' Software documentation. Unpublished.
Roads, C. 1992d. ``Granulateur.'' Software documentation. Unpublished.
Roads, C. 19921997. ``Design of a granular synthesizer.'' Unpublished design documents.
Roads, C. 1993a. ``Musical sound transformation by convolution.'' In S. Ohteru, ed.
Proceedings of the 1993 International Computer Music Conference. San Francisco:
International Computer Music Association. pp. 102109.
Roads, C. 1993b. ``Organization of Clang-tint.'' In S. Ohteru, ed. Proceedings of the
1993 International Computer Music Conference. San Francisco: International Computer
Music Association. pp. 346348.
Roads, C. 1996. The Computer Music Tutorial. Cambridge, Massachusetts: MIT Press.
Roads, C. 1997. ``Sound transformation by convolution.'' In C. Roads, S. Pope,
A. Piccialli, and G. De Poli, eds. Musical Signal Processing. Lisse: Swets & Zeitlinger.
Roads, C. 1998a. ``Espace musical : virtuel et physique.'' In F. Barriere and G. Bennett,
eds. Composition/Diusion en Musique Electroacoustique. Actes III. Academie Internationale de Musique Electroacoustique. Bourges: Editions Mnemosyne. pp. 158160.

375

References
Roads, C. 1998b. ``The Creatovox synthesizer project.'' Unpublished manuscript.
Roads, C. 1999. ``Synthese et transformation des microsons.'' Doctoral thesis. Paris:
Department of Music, University of Paris VIII.
Roads, C. 2000. ``Notes on the history of sound in space.'' Presented at the CREATE
symposium SOUND IN SPACE 2000, University of California, Santa Barbara. Revised
version forthcoming in Computer Music Journal.
Roads, C. 2001. ``Sound composition with pulsars.'' Journal of the Audio Engineering
Society 49(3).
Roads, C., and J. Alexander. 1995. Cloud Generator Manual. Distributed with the program Cloud Generator. Internet: www.create.ucsb.edu.
Roads, C., J. Kuchera-Morin, and S. Pope. 1997. ``The Creatophone sound spatializer
project.'' Unpublished research proposal.
Robindore, B. 1996a. ``Eskhate Ereuna: extending the limits of musical thoughtcomments on and by Iannis Xenakis.'' Computer Music Journal 20(4): 1116.
Robindore, B. 1996b. Program notes. Computer Music Journal compact disc. Volume
20, 1996.
Robindore, B. 1998. ``Interview with an intimate iconoclast.'' Computer Music Journal
23(3): 816.
Rocha, M. 1999. ``Les techniques granulaires dans la synthese sonore.'' Doctoral thesis.
Paris: Universite de Paris VIII.
Rodet, X. 1980. ``Time-domain formant-wave-function synthesis.'' In J. G. Simon, ed.
1980. Spoken Language Generation and Understanding. Dordrecht: D. Reidel. Reprinted
in Computer Music Journal 8(3): 914. 1984.
Rodet, X., Y. Potard, and J.-B. Barriere. 1984. ``The CHANT project: from synthesis of
the singing voice to synthesis in general.'' Computer Music Journal 8(3): 1531.
Reprinted in C. Roads, ed. 1989. The Music Machine. Cambridge, Massachusetts: MIT
Press.
Russolo, L. 1916. The Art of Noises. Translated 1986 by Barclay Brown. New York:
Pendragon Press.
Scaletti, C. 1996. ``Description of Public Organ.'' Internet: www.symbolicsound.com.
Schaeer, P. 1959. ``The interplay between music and acoustics.'' Gravensaner Blatter
14: 6169.
Schaeer, P. 1977. Traite des objets musicaux. Second edition. Paris: Editions du Seuil.
la recherche d'une musique concrete. Paris: Editions
Schaeer, P., and A. Moles. 1952. A
du Seuil.

376

References
Schaeer, P., G. Reibel, and B. Ferreyra. 1967. Trois microsillons d'exemples sonores de
G. Reibel et Beatriz Ferreyra illustrant le Traite des Objets Sonores et presentes par l'auteur. Paris: Editions du Seuil.
Schafer, R. M. 1977. The Tuning of the World. New York: Knopf.
Schafer, R., and L. Rabiner. 1973. ``Design and simulation of a speech analysis-synthesis
system based on short-time Fourier analysis.'' IEEE Transactions on Audio and Electroacoustics AU21: 165174.
Schillinger, J. 1946. The Schillinger System of Musical Composition. New York: Carl
Fischer. Reprinted 1978. New York: Da Capo Press.
Schindler, A. 1998. Eastman Csound Tutorial. Internet: www.esm.rochester.edu/onlinedocs/allan.cs.
Schnell, N. 1995. ``GRAINYgranularsynthese in Echtzeit.'' Beitrage zur Electronischen Musik 4. Fraz: Institut fur Elektronische Musik, Hochschule fur Musik und
Darstellende Kunst.
Schoenberg A. 1967. Fundamentals of Music Composition. London: Faber and Faber.
Schopenhauer, A. 1819. ``Art and the art of music.'' From The World of Will and Idea.
Translated by R. Haldane and J. Kemp, reprinted in J. Randall et al. 1946. Readings in
Philosophy. New York: Barnes and Noble. pp. 246254.
Schottstaedt, W. 2000. Common Lisp Music Manual. Online documentation. Internet:
www-ccrma.stanford.edu/CCRMA/Software/clm/clm-manual/clm.html.
Schroeder, M., and B. S. Atal. 1962. ``Generalized short-time power spectra and autocorrelation functions.'' Journal of the Acoustical Society of America 34: 16791683.
Scott, R., and S. Gerber. 1972. ``Pitch-synchronous time-compression of speech.'' Proceedings of the IEEE Conference for Speech Communication Processing. New York:
IEEE. 6365.
Serra, M.-H. 1992. ``Stochastic Composition and Stochastic Timbre: Gendy3 by lannis
Xenakis'' Perspectives of New Music 31.
Serra, M.-H. 1997. ``Introducing the phase vocoder.'' In C. Roads, S. Pope, A. Piccialli,
and G. De Poli, eds. Musical Signal Processing. Lisse: Swets and Zeitlinger. pp. 3190.
Serra, X. 1989. ``A system for sound analysis/transformation/synthesis based on a
deterministic plus stochastic decomposition.'' Stanford: Center for Computer Research
in Music and Acoustics, Department of Music, Stanford University.
Serra, X., and J. Smith. 1990. ``Spectral modeling synthesis: a sound analysis/synthesis
system based on a deterministic plus stochastic decomposition.'' Computer Music Journal 14(4): 1224.

377

References
Shensa, M. 1992. ``The discrete wavelet transform: wedding the a trous and Mallat
algorithms.'' IEEE Transactions on Signal Processing 40(10): 24642482.
Smalley, D. 1986. ``Spectro-morphology and structuring processes.'' In S. Emmerson,
ed. The Language of Electroacoustic Music. London: Macmillan. pp. 6193.
Smalley, D. 1997. ``Spectromorphology: explaining sound shapes.'' Organised Sound
2(2): 107126.
Smith, L. 1996. ``Modelling rhythm perception by continuous time-frequency analysis.''
In L. Ayers and A. Horner, eds. Proceedings of the 1996 International Computer Music
Conference. San Francisco: International Computer Music Association. pp. 392395.
Solomos, M. 1997. Program notes to Xenakis: Electronic Music. Compact disc CD 003.
Albany: Electronic Music Foundation.
Sprenger, S. 1999. ``Time and pitch scaling of audio signals.'' Internet: www.dspdimension.com/html/timepitch.html.
Steiglitz, K. 1996. A Digital Signal Processing Primer. Menlo Park: Addison-Wesley.
Steinberg. 1998. ``Cubase VST/24 audio recording.'' Karlsruhe: Steinberg Soft- and
Hardware GmbH.
Stevenson, R., and R. B. Moore. 1967. Theory of Physics. Philadelphia: W. B. Saunders.
Stockham, T. 1969. ``High-speed convolution and convolution with applications to digital ltering.'' In B. Gold and C. Rader. 1969. Digital Processing of Signals. New York:
McGraw-Hill. pp. 203232.
Stockhausen, K. 1955. ``Actualia.'' die Reihe 1. English edition. Bryn Mawr: Theodore
Presser Company. pp. 4551.
Stockhausen, K. 1957. ``. . . How time passes . . .'' die Reihe 3: 1043. English edition
translated by Cornelius Cardew. 1959. Reprinted with revisions as ``. . . wie die Zeit
vergeht . . .'' in K. Stockhausen. 1963. Texte zur elektronischen und instrumentalen
Musik. Band 1. Cologne: DuMont Schauberg: pp. 99139.
Stockhausen, K. 1961. ``Two lectures.'' die Reihe 5. English edition. Bryn Mawr: Theodore Presser Company. pp. 5982.
Stockhausen, K. 1962. ``Die Einheit der musikalischen Zeit.'' [The unity of musical
time.] Translated by E. Barkin as ``The concept of unity in electronic music.'' Perspectives of New Music 1(1): 39. Reprinted in B. Boretz and E. Cone, eds. 1972. Perspectives on Contemporary Music Theory. New York: Norton. pp. 129147. German
version published in Stockhausen (1963), pp. 211221.
Stockhausen, K. 1963. Texte zur elektronischen und instrumentalen Musik. Band 1.
Cologne: DuMont Schauberg.

378

References
Strang, G. 1994. A Friendly Guide to Wavelets. Boston: Birkhauser.
Stravinsky, I. 1936. An Autobiography. New York: W. W. Norton.
Stravinsky, I. 1947. Poetics of Music. Cambridge, Massachusetts: Harvard University
Press.
Strawn, J. 1985. ``Modelling musical transitions.'' Ph.D. dissertation. Stanford: Stanford
University Department of Music.
Stuckenschmidt, H. H. 1969. Twentieth Century Music. New York: McGraw-Hill.
Sturm, B. 1999. ``A potential G.U.T. of signal synthesis.'' Internet: www-ccrma.stanford.edu/@sturm/research.html.
Supper, M. 1997. Elektroakustische Music und Computermusic. Hofheim: Wolke Verlag.
Sussman, G., and Steele, G. 1981. ``Constraints: A language for expresssing almosthierarchical descriptions.'' A.I. Memo 502A. Articial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, Massachusetts Reprinted in Articial
Intelligence 14: 139.
Tait, C. 1995. ``Audio analysis for rhythmic structure.'' Proceedings of the 1995 International Computer Music Conference. San Francisco: International Computer Music
Association. pp. 5901.
Tempelaars, S. 1976. ``The VOSIM oscillator.'' Presented at the 1976 International
Computer Music Conference, MIT, Cambridge, Massachusetts, 2831 October.
Tempelaars, S. 1977. Sound Signal Processing. Translated by Ruth Koenig. Utrecht:
Institute of Sonology.
Tempelaars, S. 1996. Signal Processing, Speech, and Music. Lisse: Swets and Zeitlinger.
Todoro, T. 1995. ``Real-time granular morphing and spatialization of sounds with
gestural control within Max/FTS.'' In R. Bidlack, ed. Proceedings of the 1995 International Computer Music Conference. San Francisco: International Computer Music
Association. pp. 315318.
Torresani, B. 1995. Analyse continue par ondelettes. Paris: InterEditions.
Truax, B. 1984. Acoustic Communication. Norwood, New Jersey: Ablex Publishing.
Truax, B. 1986. ``Real-time granular synthesis with the DMX-1000.'' In P. Berg, ed.
1987. Proceedings of the 1986 International Computer Music Conference. San Francisco:
Computer Music Association. pp. 138145.
Truax, B. 1987. ``Real-time granulation of sampled sound.'' In S. Tipei and J. Beauchamp, eds. 1987. Proceedings of the 1987 International Computer Music Conference.
San Francisco: Computer Music Association. pp. 138145.

379

References
Truax, B. 1988. ``Real-time granular synthesis with a digital signal processing computer.'' Computer Music Journal 12(2): 1426.
Truax, B. 1990a. ``Time-shifting of sampled sound sound using a real-time granulation
technique.'' In Proceedings of the 1990 International Computer Music Conference. San
Francisco: Computer Music Association. pp. 104107.
Truax, B. 1990b. ``Composing with real-time granular sound. Perspectives of New Music
28(2): 120134.
Truax, B. 1992. ``Composition with time-shifted environmental sound.'' Leonardo Music
Journal 2(1): 3740.
Truax, B. 1994a. ``Discovering inner complexity: time-shifting and transposition with a
real-time granulation technique.'' Computer Music Journal 18(2): 3848.
Truax, B. 1994b. ``Granulation and time-shifting of sampled sound in real-time with a
quad DSP audio computer system.'' In Proceedings of the 1994 International Computer
Music Association. San Francisco: Computer Music Association. pp. 335337.
Truax, B. 1995. ``Sound in context: soundscape research and composition at Simon
Fraser University.'' In Proceedings of the 1995 International Computer Music Conference. San Francisco: International Computer Music Association. pp. 14.
Truax, B. 1996a. ``Time-stretching of hyper-resonated sound using a real-time granulation technique. In Proceedings of the 1996 International Computer Music Conference.
San Francisco: International Computer Music Association. pp. 491492.
Truax, B. 1996b. ``Soundscape, acoustic communication, and environmental sound
composition.'' Contemporary Music Review 15(1): 4965.
Tyndall, J. 1875. Sound. Akron: Werner.
Ungeheuer, E. 1992. Wie de elektronische Musik ``erfunden'' wurde . . . Mainz: B.
Schott's Sohne.
Vaggione, H. 1984. ``The making of Octuor.'' Computer Music Journal 8(2): 4854.
Reprinted in C. Roads, ed. 1989. The Music Machine. Cambridge, Massachusetts: The
MIT Press. pp. 149155.
Vaggione, H. 1996a. ``Autour de l'approche electroacoustique : situations, perspectives.''
In G. Bennett, C. Clozier, S. Hanson, C. Roads, and H. Vaggione, eds. Esthetique et
Musique Electroacoustique. Bourges: Editions Mnemosyne. pp. 101108.
Vaggione, H. 1996b. ``Articulating micro-time.'' Computer Music Journal 20(2): 3338.
Vaggione, H. 1996c. ``Vers une approche transformationelle en CAO.'' Actes des JIM
1996. Les cahiers du GREYC. Caen: CNRS-Universite de Caen. Internet: www.
ircam.fr/equipes/repmus/jim96/actes/Vaggione/VaggioneTEXT.html.

380

References
Vaggione, H. 1997. ``Singularite de la musique et analyse: l'espace d'intersection.'' In F.
Barriere and C. Clozier, eds. Analyse en Musique Electroacoustique. Bourges: Editions:
Mnemosyne. pp. 7481.
Vaggione, H. 1998. ``Transformations morphologiques: quelques exemples.'' Actes des
JIM 1998.
Vaggione, H. 1999. Personal communication.
Varese, E. 1940. ``Organized sound for the sound lm.'' Quoted in Miller (1945).
Varese, E. 1962. From a lecture given at Yale University. Reprinted in Varese (1971).
Varese, E. 1971. ``The liberation of sound.'' In C. Boretz and E. Cone, eds. Perspectives
on American Composers. New York: Norton. pp. 2634.
Vercoe, B. 1993. Csound: A Manual for the Audio Processing System and Supporting
Programs with Tutorials. Cambridge, Massachusetts: MIT Media Laboratory.
Vetterli, M., and C. Herley. 1992. ``Wavelets and lter banks: theory and design.'' IEEE
Transactions on Signal Processing ASSP40(9): 22072234.
Walker, B., and K. Fitz. 1992. Lemur Manual. Urbana: CERL Sound Group, University of Illinois.
Wannamaker, R., and E. Vrscay. 1997. ``Fractal wavelet compression of audio signals.''
Journal of the Audio Engineering Society 45(7/8): 540553.
Weare, C. 1997. Personal communication.
Weidenaar, R. 1989. ``The Telharmonium: A History of the First Music Synthesizer,
18931918.'' Ph.D. dissertation. New York: New York University.
Weidenaar, R. 1995. Magic Music from the Telharmonium. Metuchen: Scarecrow Press.
Weinberg, S. 1983. The Discovery of Subatomic Particles. New York: W.H. Freeman.
Wenger, E., and E. Spiegel. 1999. MetaSynth Manual. San Francisco: U&I Software.
Whitehouse, D. 1999. ``Haunting sound from the dawn of time.'' British Broadcasting Company. 23 September 1999. Internet: news.bbc.co.uk/hi/english/sci/tech/
newsid_454000/454594.stm
Whiteld, J. 1978. ``The neural code.'' In E. Carterette and M. Friedman, eds. 1978.
Handbook of Perception, vol. IV, Hearing. Orlando: Academic.
Wickerhauser, V. 1989. ``Acoustic signal compression with wave packets.'' New Haven:
Department of Mathematics, Yale University.
Wiener, N. 1964. I Am a Mathematician. Cambridge, Massachusetts: MIT Press.

381

References
Wiener, N. 1964. ``Spatial-temporal continuity, quantum theory, and music.'' In M.
Capek, ed. 1975. The Concepts of Space and Time. Boston: D. Reidel.
Winckel, F. 1967. Music, Sound, and Sensation. New York: Dover Publications.
Winham, G. 1966. The Reference Manual for Music 4B. Princeton: Princeton University
Music Department.
Winham, G., and K. Steiglitz. 1970. ``Input generators for digital sound synthesis.''
Journal of the Acoustical Society of America 47(2, Part 2): 665666.
Winston, P. 1984. Articial Intelligence. Second edition. Reading, Massachusetts:
Addison-Wesley.
Wishart, T. 1994. Audible Design. York: Orpheus the Pantomime.
Wishart, T. 1996. Program notes for Tongues of Fire. Internet: www.aec.at/prix/1995/
E95gnM-tongues.html.
Xenakis, I. 1960. ``Elements of stochastic music.'' Gravensaner Blatter 18: 84105.
Xenakis, I. 1971. Formalized Music. Bloomington: Indiana University Press.
Xenakis, I. 1989. ``Concerning time.'' Perspectives of New Music 27(1): 8492.
Reprinted in French as ``Sur le temps'' in I. Xenakis, 1994, Keleutha. Paris: L'Arche.
Xenakis, I. 1992. Formalized Music. Revised edition. New York: Pendragon Press.
Yermakov, A. 1999. SoundFront documentation. Internet: www.kagi.com

This page intentionally left blank

The Cloud Generator Program

Features of Cloud Generator

Start Time and Duration
Density
Bandlimits
Cloud Type
Stereo Location
Grain Duration
Selection Order
Waveform
Text/AIFF Output
This appendix describes Cloud Generator (CG), an interactive MacOS program for
granular synthesis and sound le granulation (gure A.1). Since its development, I have
taught hundreds of musicians how to operate this easy-to-use software. It appears that
CG will soon be obsolete, owing to changes in the underlying operating system. On an
appropriate platform, however, the program is still useful for teaching and for experimentation. (See www.create.ucsb.edu for information on our newer program for particle
synthesis, PulsarGenerator.)
John Alexander and I wrote Cloud Generator in the C programming language in
1995. At the time, John Alexander was my student at Les Ateliers UPIC (now CCMIX)
in Paris. We based CG on a suite of my 19881992 programs for granular synthesis. (See
description in chapter 3.) These programs were not easy to use, since they required users
to operate within environment of a C compiler.

384

Appendix A

A.1 Screen image of the Cloud Generator program.

CG merged the code from several of these programs into a single stand-alone application. It unied them under a common interface. Our goal was to create a tool for experimentation, teaching, and composition with granular techniques.
We created CG before Power PC microprocessors were generally available for Apple
computers. Due to changes in MacOS, it would require considerable eort to update the
program for newer processors. Nonetheless, the non-oating-point (NFPU) version runs
in emulation mode on any Power Macintosh computer up through the G3/450 MHz
running MacOS System 8.6. The performance is acceptable; many clouds compute in
faster than real time.
Since we developed CG, I have continued to extend my earlier programs for purposes
of experimentation. Since these programs do not depend on interaction with a graphical
user interface, they can more easily be programmed to perform operations not in CG.
Each program in the suite performs a particular type of granular synthesis or granular
transformation. Since 1998, Alberto de Campo and I moved most of the functionality of
CG to the Creatovox instrument, which chapter 5 documents.

385

The Cloud Generator Program

Features of Cloud Generator

We designed the front panel of CG to look like a laboratory signal generator. The console contains all the parameters that a user needs to control granular synthesis.
CG emits two types of grains: synthetic (which repeat a xed waveform), and granulated (which extract waveforms from dierent locations in a sound le). The synthetic
grain mode oers a choice of preset grain waveforms, such as sine, square, and sawtooth. Users can also extract any waveform stored in a sound le, including sampled
waveforms. Finally, the program provides a graphical editor for drawing grain waveforms manually. High-resolution interpolating oscillators generate the synthetic grains.
In granulation mode, CG accepts a stereo sound le as its input. After loading this le
into its memory space, CG extracts sound particles from the soundle according to
random, statistical, or deterministic progression algorithms (described later).
The output of CG is either a synchronous cloud or an asynchronous cloud. In a synchronous cloud, the grains follow one another after a xed time interval. In an asycnhronous cloud, the grains scatter randomly in time.
As it is calculating, CG displays the formation of the cloud as a plot of time versus
frequency. It produces a stereo soundle output. If desired, it can also produce a textle,
listing the numerical properties of each grain in a cloud. Plotting programs can can read
this text. In one, exceptional, case, the composer Mark-David Hosale used CG to produce texts that he transcribed into an instrumental composition.
The next sections explain the meaning of the parameters displayed on the console.

Start Time and Duration

The start time and duration parameters set the temporal boundaries of a cloud, indicated
in seconds. The usual start time of a cloud is 0. A positive start time, such as 2.0, creates
a sound le containing silence for two seconds, followed by the cloud. This feature could
be useful in precisely synchronizing multiple soundles in an editing and mixing program. If the output of the program is text, this feature could be used to order multiple
clouds in a plotting program.
The duration of a cloud can range from 0.000001 to an upper bound of 47.0 seconds,
given sucient memory. The reason for the upper bound on cloud duration is that the
synthesis data is calculated in random-access memory (RAM) before being written to
disk. The program preallocates a buer large enough to accomodate 47 seconds of stereo
audio for a synthetic cloud.

Density
Grain density is a function of grains per second. A synchronous cloud generates metrical
rhythmic sequences when grain densities are in the range of between two and twenty
grains per second. When the initial density is not the same as the nal density, this
creates an acceleration or deceleration eect.

386

Appendix A
For asynchronous clouds, a variation between the initial and nal density produces an
irregular acceleration or deceleration. To create continuous noise sounds, one sets the
density to above one hundred grains per second.

Bandlimits
Bandlimits describe the frequencies of the four corners of a trapezoidal region inscribed
on the frequency-versus-time plane. The grains scatter within the bounds of this region.
To create a line of grains at a xed frequency, one sets all the bandlimits to the same
frequency, for example, 440 Hz. To scatter grains in a rectangular region between 80 Hz
and 300 Hz, one sets both high bandlimits to 300, and both low bandlimits to 80. For
a cloud that converges to 300 Hz, one would set both nal bandlimits to 300. (See
gure 3.6.)
Bandlimits apply only in the case of a cloud lled by synthetic grains, where the
selected waveform is of the type ``Synthetic.'' They have no meaning when the cloud is a
granulated sound le.

Cloud Type
This parameter species the temporal distribution of the grains, either synchronous or
asynchronous. Synchronous generation means that one grain follows another in series,
with the distance between the grain attacks determined by the density parameter. For
example, a synchronous cloud with a density of ve grains per sec creates quintuplets at
a tempo of M.M. 60. The spacing between grain attacks is 200 ms. Asynchronous clouds
scatter grains at random time points within the specied boundaries of the cloud, at the
specied density. Asynchronous clouds produce aleatoric and explosive eects.

Stereo Location
The stereo location potentiometers provide an intuitive interface for the initial and nal
position of the sound cloud. By clicking on either the initial or nal Random box the
grain scatters to a random position in stereo space. This randomization articulates
the granular texture of a cloud. It enhances the impression that the cloud is a threedimensional object in space.

Grain Duration
Grain duration has a strong inuence on the timbre of the resulting cloud. Grain duration is a value in seconds. Thus a value of 0.1 equals 100 ms. The initial grain duration
applies at the beginning of the cloud. The nal grain grain duration applies at the end. If
the Random box is checked, the grain duration parameters switch from ``Initial'' to
``Minimum'' and from ``Final'' to ``Maximum.'' In this case, the grain duration is a
random function between the specied minimum and maximum values.

387

The Cloud Generator Program

Selection Order
The Selection Order parameter applies only to granulated clouds. It determines the
order in which grains are selected from the input sound le. Three options present
themselves:
1. Randomselect input grains from random points in the input sound le.
2. Statistical evolutionselect input grains in a more-or-less left-to-right order. That is,
at the beginning of the cloud there is a high probability that grains derive from the
beginning of the input le; at the end of the cloud there is a high probability that
grains derive from the end of the input le.
3. Deterministic progressionselect input grains in a strictly left-to-right order.
Option 1 results in a scrambled version of the input. Option 3 preserves the temporal
order of the original (we use this mode when time-shrinking or time-stretching the input
sound le). Option 2 is a blend of options 1 and 3.

Waveform
The waveform parameter has several controls. The rst is a choice between Synthetic or
Granulated source waveforms. The name of the chosen waveform appears in the small
box below the buttons.
1

In a synthetic cloud, an oscillator reads a wavetable containing one period of a waveform at a frequency within the high and low bandlimit boundaries.

In a granulated cloud, the grains are extracted from a stereo input sound le of any
length. The extracted grains are scattered in time without pitch transposition. The
bandlimit parameters are not operational in a granulated cloud.

The Select button brings up a dialog box that depends on whether the cloud is synthetic or granulated. If it is synthetic, the dialog box displays ve choices:
1. Sine
2. Sawtooth
3. Square
4. User-drawnone can freely draw in the waveform editor
5. Imported from a sound le. The imported le must be exactly 2048 samples (46 ms)
in length. If the waveform is extracted from a sampled sound le, this waveform will
repeat at its extracted frequency only if the cloud bandlimits are set to 21 Hz. Any
frequency above this transposes the waveform
If the cloud is granulated, the Select button displays an Open File dialog box. The user
is invited to select a stereo sound le to granulate.

388

Appendix A

Text/AIFF Output
This option lets one save the cloud data in the form of numerical text. This text could be
read by a plotting program to make a graphical score. Clicking on AIFF creates a sound
le as the output of the program's calculations. This is the normal mode.

Sound Examples on the CD

This recording documents the history of synthesis with microsonic particles and presents
artistic excerpts of music compositions realized with these techniques. It also serves as an
audio notebook of research experiments in this domain.
Tracks 1 through 3 present the historical examples. Tracks 4 through 13 present musical excerpts. Tracks 14 through 44 present an audio notebook of research experiments
in particle synthesis. Tracks 45 through 68 present an audio notebook of experiments in
microsonic sound transformation.

Historical Examples
1. Three excerpts of Analogique A et B for string orchestra and analog granular synthesis by Iannis Xenakis. The piece was composed in 1958 and 1959. The composition
opposes part A for string orchestra, and part B for tape. (a) Tape part alone. (b) Tape
with instruments in a busy section. (c) Tape with instruments in a sparse section.
Reproduced with the permission of Iannis Xenakis and Editions Salabert, 22 rue
Chauchat, 75009 Paris, France, www.salabert.fr. [1:03]
2. Klang-1 (Curtis Roads, December 1974), rst granular synthesis experiment by computer. The recording is distorted, due to technical factors. [0:48]
3. Excerpt of Prototype (Curtis Roads, April 1975), rst study in automated granular
synthesis. [0:26]

Sound Particles in Music Compositions

4. Excerpt of nscor (Curtis Roads, 1980). This excerpt [4:305:40] features a granular
explosion followed by a passage of granular noise clouds in dierent frequency
bands. [1:14]
5. Excerpt of Field (Curtis Roads, 1981). This excerpt [3:374:02] features a granular
explosion. [0:28]

390

Appendix B
6. Excerpt of Organic (Curtis Roads, 1994). This excerpt [1:552:49] opens with an
asynchronous pulsar cloud. All of the beating and pulsating sounds are pulsars.
[0:58]
7. Excerpt of Half-life (Curtis Roads, 1999). This excerpt [0:001:04] from the rst
movement, entitled Sonal Atoms, consists of a pulsar stream that has been granulated and then extensively edited on a microsonic time scale. [1:09]
8. Excerpt of Tenth vortex (Curtis Roads, 2000). This excerpt [0:001:29] consists of a
constantly evolving cloud of grains produced by my constant-Q granulation program applied to a stream of synthetic pulsars. [1:28]
9. Excerpt of Eleventh vortex (Curtis Roads, 2001). The opening of this excerpt [0:00
1:30] features multiple synchronous pulsar streams, each at a separate tempo. The
thick stream of particles thins into a series of discrete particles about 40 seconds into
the piece. [1:05]
10. Excerpt of Sculptor (Curtis Roads, 2001). The source material of Sculptor was a
driving percussion piece by the group Tortoise, sent to me for processing in August
2000 by John McEntire. I granulated and ltered the material by means of the
constant-Q granulator, which disintegrated the beating drums into a torrent of
sound particles. Beginning with this whirling sound mass, I articulated the internal
morphologies and implied causalities within the current of particles. This meant
shaping the river of particle densities, squeezing and stretching the amplitudes of
individual particles and particle clouds, carving connected and disconnected frequency zones, and twisting the spatial ow. Over months of intermittent editing on
dierent time scales, I was able to sculpt this material into its current form. The
composition was completed in July 2001 in Santa Barbara. [1:31]
11. Excerpt of Agon (Horacio Vaggione, 1998). Agon is one of a series of works by
Horacio Vaggione that explore the concept of composition on multiple time scales.
The original sound sources were acoustic percussion instruments, which have been
broken down into microsonic particles and reassembled by the composer in a technique of micromontage. [1:08]
12. Excerpt of Life in the Universe (Ken Fields, 1997). This excerpt [0:001:36] features
the granulation of the synthetic voice of physicist Steven Hawking. It appeared on a
CD-ROM entitled Steven Hawking Life in the Universe that was marketed by MetaCreations. [1:42]
13. Three excerpts from the electronic part of Paleo for double bass and CD. (JoAnn
Kuchera-Morin, 2000) demonstrating the use the phase vocoder to time scale
sounds. (a) The rst 45 seconds of the computer-generated portion of the work. The
sound consists of processed double bass and dijeridu string Csound instruments. No
time scaling. (b) From the middle section of the work, a passage time-stretched by a
factor of 7.9. (c) From the middle section of the work, a passage time-shrunk by a
factor of 0.5. [2:36]

391

Sound Examples on the CD

Experiments in Particle Synthesis

Synthetic Granular Synthesis
14. A melodic stream of sinusoidal grains, increasing regularly in density over 40 seconds. The frequency band within which the grains are scattered opens up to a high
frequency of 5000 Hz and a low frequency of 80 Hz. The pitch is random within this
range. The grain duration is also diminishing over time, from 310 ms to 1 ms, resulting in a short pop at the end. Reverberation added. [0:42]
15. Synchronous granular cloud with a wide bandwidth at a constant density of fteen
grains per second, followed by an asynchronous cloud with the same specications.
Reverberation added. [0:30]
16. In a synchronous granular cloud, the grain duration and density parameters may be
a stronger determinant of pitch than the frequency of the waveform inside the grain.
This example presents a synchronous cloud of 10-ms sinusoidal grains with a nominal frequency of 200 Hz. The pitch changes as the density increases from 100 to 200
grains per second. In eect, the onset time of the grains becomes the perceived fundamental pitch. [0:12]
17. Dense cloud made up of sinusoidal grains. The initial density of 600 grains per second, diminishing to 1. The initial frequency bandwidth is 13 kHz down to 40 Hz,
and the nal frequency bandwidth is centered at 200 Hz; hence the narrowing frequency band. The duration of the grains is random between 12 ms and 1 ms. [0:11]
18. Sparse, broadband cloud of very short grains, with durations between 1 ms and 500
msec. Reverberation added. [0:11]
19. Metric stream of grains, 30 ms in duration, in a looped pattern. Reverberation
added. [0:25]
20. Real-time granular synthesis performed on the Creatovox synthesizer. A collage of
Creatovox sounds originally performed by composer Bebe Barron in Studio Varese,
Center for Research in Electronic Art Technology (CREATE), University of California, Santa Barbara, July 2000. Ms. Barron's musical assistant was Alberto de
Campo. The method of synthesis was synthetic granular synthesis. Reverberation
was added at the time of recording. This collage was assembled by Curtis Roads in
August 2001. [1:18]

Glisson Synthesis
21. Isolated glissons. (a) Two 250 ms glissons, one sweeping from 1000 Hz to 500 Hz,
the other in the opposite direction. (b) Two 56 ms glissons, sweeping from 55 Hz to
1400 Hz and the other in the opposite direction. (c) Two sets of ``bird call'' glissons.
[0:11]

392

Appendix B
22. Trains of 6 ms glissons with small frequency deviations centered at 1450 Hz and 400
Hz respectively. [0:07]
23. Seven glisson clouds. (a) High-density cloud, wide frequency sweep. (b) Long glissons, high density cloud, converging on 550 Hz. (c) Long glissons converging on
1100 Hz. (d) High-density glisson cloud sweeping from 6 to 2000 Hz. (e) Long glissons. (f ) Short glissons. (g) High-frequency glissons. [1:46]

Grainlet Synthesis
Experiments in grainlet synthesis, 19961998. These experiments demonstrate linkages
between grainlet frequency and grainlet duration. All experiments realize asynchronous
(non-metric) clouds. I have added reverberation and some editing eects, primarily particle replication.
24. Sparse cloud in which high frequency grainlets are short, while low-frequency
grainlets are long. [0:10]
25. Four clouds. (a) One cycle (fundamental period) per grainlet. (2) Four cycles per
grainlet. (3) Eight cycles per grainlet. (d) Sixteen cycles per grainlet. [0:33]
26. Medium-density grainlet cloud, with short high frequencies and long low frequencies. [0:23]
27. A cloud with a frequency pole at 440 Hz, making grainlets around that frequency
much longer than low frequency and high frequency grainlets. [0:13]
28. Three high-density (90 particles per second) grainlet clouds. (a) Two cycles per
grainlet. (b) Four cycles per grainlet. (c) Twenty cycles per grainlet. [0:20]
29. Six high-density grainlet experiments.

[1:06]

Trainlet Synthesis
Experiments in trainlet synthesis, 19992000.
30. Individual trainlets. (a) Gbuzz test demonstrating increasing ``lowest harmonic.''
(b) 40 Hz trainlet with 16 harmonics and chroma 1.0. (c) 600 Hz trainlet with
chroma 1.0. (d) 600 Hz trainlet with chroma 0.5. [0:34]
31. Sparse trainlet phrases. [1:00]
32. Dense trainlet clouds.

[1:02]

Pulsar Synthesis
Experiments in pulsar synthesis, 19912001.
33. Cosmic pulsar from the neutron star PSR 0329+54. [0:07]

393

Sound Examples on the CD

34. Pulsar variations. (a) Constant fundamental frequency, three variable formant frequencies. Each formant has an independent trajectory in space. (b) Variable fundamental frequency, three constant formant frequencies (low, mid, high frequency).
Each formant has an independent trajectory in space. [0:49]
35. Pulsar rhythm versus tone. (a) Pulsar rhythm. (b) Pulsar tone, with the same formant envelope as the previous rhythm example. [0:53]
36. Rhythmic burst masking examples, 8 Hz pulsar rhythm. One second of silence separates each example. (a) No burst masking. (b) 2:1 burst ratio. (c) 4:1 burst ratio. (d)
5:1 burst ratio. (e) 5:5 burst ratio. [1:03]
37. Tone burst masking, subharmonics, and long tone pulses. 260 Hz pulsar rhythm
with the same formant structure as the previous example. The only dierence is the
fundamental frequency. One second of silence separates each example. (a) No burst
masking makes a constant tone. (b) 2:1 burst ratio, subharmonic eect. (c) 4:1 burst
ratio, subharmonic eect. (d) 5:1 burst ratio, subharmonic eect. (e) 5:5 burst ratio,
resulting in a slow amplitude modulation eect. (f ) 5:10 burst ratio, resulting in
discrete pulses. (g) 14:N burst ratio, where N is increasing, resulting in long
pulses. [1:06]
38. Stochastic masking in a low-frequency 40 Hz tone. Notice the intermittency in
the pulsar stream. The constantly-varying formants are independent of this intermittency. [0:22]
39. Eight short pulsar trains demonstrating a variety of pulsar sounds, from strong formant streams to rhythms and noise. Some examples have been edited and processed
with reverberation. [3:36]

Graphic and Sonographic Synthesis of Microsound

40. Synthesis of microsound in Phonogramme. This example presents a sparse scattering
of microarcs, followed by a more dense cloud of microarcs. [0:14]
41. Two examples of synthesis of microsound in MetaSynth. A complex cloud painted
with MetaSynth's spray jet brush. Each particle is played by a sample playback oscillator. Both examples correspond to the same image score, but are edited afterward
so that they have dierent accent points. (a) Sample playback with mallet instrument samples. (b) Sample playback with percussion instruments. [0:28]

Synthesis by Transient Drawing and Particle Cloning

42. Synthesis by transient drawing. (a) 4.75 sec of transient waveforms drawn in a sound
editor with a stylus tool (Alchemy by Passport Designs). (b, c, d) Beats and tones
made by extracting particles from (a) and replicating and transposing them. (e) Six
tones made from a single particle, replicated and transposed. [0:24]

394

Appendix B
43. Transient transformation, in which a single transient event is extracted from a longer
sound and then reshaped into dierent sounds through editing operations. Each example is separated by two seconds of silence. (a) The original source, a noisy onesecond sound recorded by a hydrophone. (b) 125-ms extraction from (a). (c) The
same event as (b), pitch-shifted up an octave, cloned ten times, and shaped with an
exponential decay envelope. (d) The same event as (b), strongly ltered to create a
colored tone. (e) The same event as (b) reversed and re-enveloped into a percussive
sound. (f ) The same event as (e), pitch-shifted and replicated to form two pitched
tones. [0:18]

Physical Models of Particles

44. Examples of Perry Cook's maracas model as realized by J. Bumgartner's SYD program. Each sound object is like a single shake of a maraca. (a) Resonance frequency
100 Hz. (b) Resonance frequency 700 Hz. (c) Resonance frequency 3200 Hz. This is
probably closest to the frequency of real maracas. (d) Resonance frequency 3200 Hz,
but with a low probability of bean collisions. This models the case of a small number
of beans in the maraca. (e) Strong resonance frequency 80 Hz. (f ) Strong resonance
frequency 600 Hz. [0:24]

Experiments in Sound Transformation with Particles

Micromontage
45. Result of micromontage in a graphical sound editing and mixing program, corresponding to gure 5.1. [0:04]

Granulation
46. Basic granulations. (a) Original spoken Italian utterance. (b) ``Robotic'' granulation
transformation by layering several copies slightly delayed. (c) ``Thickening'' with
increasing delay. (c) ``Scrambling'' with large grains. (d) Scrambling with mediumsized grains at medium density. (e) Scrambling with tiny grains and high density,
going to long grains and low density. (f ) Granulation with pitch shifting for a chorus
eect. (g) ``Graverberation'' (granular reverberation) eects realized with the SampleHose by Michael Norris. [1:34]
47. Timbral enrichment by particle envelope manipulation. (a) Edgard Varese speaking.
(b) Change in formant location caused by many overlapping grains. (c) Short grains
make the speech noisy. [0:44]

395

Sound Examples on the CD

Pitch-Shifting on a Micro Time Scale

48. Experiments with pitch-shifting on a micro time scale. (a) Italian spoken phrase cut
into 150 ms grains, each of which is pitch-shifted up to 40% up or down. (b) Same
phrase with 50 ms grains and broader range of pitch-shifting. (c) Phrase with ten
times grain overlap within a narrow range up and down. (d) Phrase with ten times
overlap within a broad range up and down, creating a chorus eect. [0:25]

Granular Time Stretching and Shrinking

49. Example of granular time-stretching using Cloud Generator. (a) Statement by Iannis
Xenakis on ``quantum sound synthesis,'' recorded 1989 by Christian Clozier. (b)
Time stretched by a factor of two, with grain density increasing from 10 to 50 grains
per second while the grain duration decreases from 100 ms to 50 ms. (c) Time shrunk
to one-half the original duration. The grain durations are constant at 75 ms and the
density is 50 grains per second. In both examples, the grains are randomly scattered
within the stereo eld. [0:36]

Granular Formant Shifting

50. Granular formant shifting. The fundamental pitch and the duration of the sound
remain constant. (a) Original Italian phrase. (b) 3 semitones formant shift. (c) 12
semitones formant shift. (d) 3 semitones formant shift. (e) 12 semitones formant
shift. [0:27]

Filtering on a Micro Time Scale

51. Experiment with constant-Q lters on grains. (a) Excerpt of Hopkinson Smith
playing the lute composition Gagliardi 5 from Libro Primo d'Insta Volatura di Lauto
by G. G. Kapsberger (15801651) recorded on Astree E 8553. (b) Constant-Q granulations of (a), including time scrambling of the grains. Realized in real-time with
SuperCollider. [1:58]
52. Experiment with constant-Q lters on grains. (a) Excerpt of percussion track sent to
me by John McEntire, July 2000. (b) Excerpt of Tsound, a study made by constantQ granulation of (a) with additional editing on a micro time scale. [2:44]

Dynamics Processing on a Micro Time Scale

53. Granular dynamics processing. (a) Original Italian phrase. (b) Compression and
expansion of high frequency band above 4000 Hz. (c) Processed with mid-frequency
compansion. (d) Processed with extreme compansion focused on a narrow frequency
band (452 Hz and its harmonics). [0:14]

396

Appendix B

Waveset and Wavecycle Distortions

54. Waveset transformations, realized with the SoundMagic FX package by Michael
Norris, a plugin set for SoundMaker. (a) Original Italian phrase. (b) Waveset harmonic distortion. (c) Time-varying waveset omission. (d) Waveset reversal. The
waveset size is 60. (e) Waveset time-stretching by a factor of 7.6. [0:58]

Convolution of Microsounds
Experiments in convolving sound particles with other sounds.
55. Granular convolution examples. (a) Synchronous grain cloud of grains lasting 7
seconds. The grain durations are between 1 and 9 ms. (b) Snare drum rim shot. (c)
Convolution of (a) and (b). (d) Convolution of (b) with an asynchronous grain cloud
with increasing density up to 90 grains per second. (e) 12-second synchronous cloud
of sinusoidal grains at a density of 7 grains per second. (f ) Swirling electronic texture
lasting 7 seconds. (g) Convolution of (f ) and (g), which extends the duration to 19
seconds. [0:16]
56. Convolutions with a dense cloud of sonic grains to create quasi-reverberation eects.
Convolution of Italian spoken phrase ``Lezione numero undice, l'ora'' with a dense
cloud of Gaussian grains. The convolution window is of the Hamming type. [1:25]
57. Pulsar convolution example. (a) Italian word ``qui.'' (b) Pulsar train. (c) Convolution of (a) and (b). [0:26]

Sound Transformation with Windowed Spectrum Operations

58. Phase vocoder operations: pitch-shifting with constant duration. (a) Edgard Varese
speaking. (b) Pitch shifting up ve semitones. (c) Pitch-shifting down ve semitones. [0:54]
59. Phase vocoder operations: time stretching and shrinking with constant pitch. (a)
Edgard Varese speaking. (b) Time stretching by a factor of two. (c) Time shrinking
by a factor of one-half. [0:37]
60. Exotic spectrum transformations that manipulate the Gabor matrix. These examples
were realized with the SoundMagic and SoundFront plugins to the SoundMaker
editor. (a) Original Italian phrase. (b) Spectral magnet aligned to a pentatonic scale.
(c) Spectral blurring with a 200 ms frame. (d) Spectral lter sweeping up. (e) Spectral
randomizer. (f ) Two examples of spectral tracing: the softest 1% of the spectrum
followed by the loudest 1%, both normalized. (g) Spectral shift, upward sweep. (h)
Spectral mirror, inversion around 1 kHz. [0:47]
61. A classic vocoder transformation. (a) Speaking voice. (b) Fragment of orchestra
music (Sinfonia sopra una canzone d'amore by Nino Rota, 1947, Fonit Cetra NFCD
2020). (c) Vocoding of (a) and (b), where (a) is the lterbank modulator and (b) is
the carrier. The eect was realized with Arboretum HyperPrism. [0:52]

397

Sound Examples on the CD

Sonographic Transformations
62. Sonogram ltering. (a) Voice of Edgard Varese. (b) Voice ltered by inscribing multiple regions on a sonogram and reducing them by 24 dB. [0:26]

Transformations Based on the Tracking Phase Vocoder

63. Tracking phase vocoder (TPV) manipulations 1, manipulations based on track
amplitudes. (a) Spoken French phrase. (b) Resynthesized from a TPV analysis. (c)
Resynthesis after deletion of low-level tracks less than 80 dB down. (d) Resynthesis
after deletion of low-level tracks less than 30 dB down. (e) Resynthesis with deletion of all tracks greater than 100 ms in duration. (f ) Resynthesis with deletion of all
tracks less than 100 ms in duration. [0:27]
64. TPV transformations 2, spectrum shifting and multiplying. (a) Original Italian
phrase. (b) Spectrum frequency shifted by 200 Hz. (c) Spectrum multiplied by 0.8.
[0:18]
65. TPV transformations 3, disintegration. (a) Original Italian phrase. (b) Granular
disintegration caused by deleting more and more time-frequency tracks (up to 85%
by the end). [0:12]
66. TPV transformations 4, morphing sounds. (a) Flute morphing into a clarinet. (b)
Flute morphing into a cello. Examples courtesy of Kelly Fitz. [0:15]

The Vector Oscillator Transform

67. (a) Spoken Italian phrase. (b) VOT resynthesis with 100 Hz fundamental. (c) VOT
resynthesis with 60 Hz fundamental. (d) Time stretched version. [0:16]

Wavelet-Based Transformations
68. Comb wavelet separation of harmonic and inharmonic components. (a) Violin note.
(b) Harmonic part of violin tone. (c) Inharmonic part of violin tone. (d) Low piano
tone. (e) Harmonic part of piano tone. (f ) Inharmonic part of piano tone. [0:19]

This page intentionally left blank

Name Index

Alexander, J., 111, 383

Arb, D., ix, 113, 298
Aristotle, 53
Arroyo, R., 319
Babbitt, M., 78, 81, 350
Bach, J. S., 11, 268
Backus, J., 81
Barlow, C., 116
Barrett, N., 322
Barron, B., 195
Beekman, I., 512
Beethoven, L., 11
Behles, G., 115
Bencina, R., 114
Bertezki, A., 115
Bertrand, R., 68
Blum, T., x, 323
Boethius, 53
Boulez, P., 71
Boyle, R., 53
Brummer, L., 319
Brun, H., 30, 348
Cage, J., 13, 233, 329
Cardew, C., 72
Cascone, K., 323
Cavaliere, S., x, 112
Chen, C.-W., 195
Cheng, C., x
Chion, M., 82

Chowning, J., 175

Chrysippus, 52
Cloud, D., 305
Clozier, C., 231
Cook, P., 97, 174
Cooley, J., 244
Cope, D., 346
Cottrell, L., 109
Cowell, H., 556, 329
Dahlhaus, C., 332
Dalton, J., 52
Daubechies, I., 284
Debussy, C., 12
de Campo, A., ix, 97, 122, 143, 195, 223, 384
Delprat, N., 113
Descartes, R., 51
de Reydellet, J., x
De Tintis, R., 114
Di Scipio, A., 322
Dolson, M., 247, 318
Eastman, G., 57
Eckel, G., 269
Edison, T., 57
Eimert, H., 61, 71
Einstein, A., 34, 54
Eloy, J.-C., 11
Epicurus, 51
Erbe, T., x, 114
Evangelista, G., x, 112

400

Name Index
Ferrari, L., 232, 312
Fields, K., 321
Fokker, A., 81
Fourier, J., 243
Fritton, W., 56
Fry, C., 190
Gabor, D., vii, 27, 54, 5760, 62, 84, 86,
250, 2956, 300, 302, 349
Galileo, G., 53
Gassendi, P., 51
Goeyvaerts, K., 71
Gredinger, P., 71
Hawking, S., 321
Heisenberg, W., 250
Helmholtz, H., 17
Helmuth, M., 113
Heraclitus, 14
Hiller, L., 337
Hinton, M., 112
Howard, E., 116
Huygens, C., 50
Kandinsky, W., 328
Kavina, L., 82
Keagi, W., 69
Keller, D., 97, 115, 322
Kelvin, L., 243
Koenig, G. M., 29, 68, 71, 812, 131, 138
Koenigsberg, C., 81
Krenek, E., 71
Kronland-Martinet, R., 288
Kuchera-Morin, J., ix, 318
Kussmaul, C., x
La Barbara, J., 321
Lansky, P., 186, 317
Lavoisier, A., 192
Le Duc, J.-C., 231
Leibig, B., 109
Lesbros, V., 15960
Ligeti, G., 15, 83, 332
Lippe, C., 113
Lopez, S., 114

Lucchese, L., x
Lucretius, 51
Mallat, S., 284
Mart, F., 114
Mauchly, W., 115
Maxwell, J., 50
McCartney, J., 113, 193
McLaren, N., 157
Mersenne, M., 51, 53
Meyer-Eppler, W., 24, 62, 69, 300
Miranda, E., 116
Moles, A., 63
Morlet, J., 284
Mozart, W., 11, 168, 345
Nelson, G., 112
Nelson, J., 321
Newton, I., 49, 63
Norris, M., 116
Omsky, J., x
Pape, G., ix, 319
Parmegiani, B., 15
Penderecki, W., 83
Piccialli, A., x, 112
Pope, S., ix, 122
Poullin, J., 302
Pound, E., 556
Pousseur, H., 316
Rayleigh, L., 173
Reihn, R., 317
Resina, E., 114
Risset, J.-C., viii, ix, 2989
Roads, M., x
Robindore, B., x, 320
Rocha, M., 321
Rockmore, C., 82
Roget, P., 56
Rolfe, C., 115
Roy, P., x
Roy, S., 318
Russolo, L., 233, 327

401

Name Index
Scaletti, C., 323
Schaeer, P., 17, 20, 61, 181, 302, 328
Scherchen, H., 62
Schnell, N., 114
Schoeller, P., 319
Schoenberg, A., 328
Scott, D., 318
Serra, X., 278
Sery, D., x
Seurat, G., 171
Sica, G., x, 319
Smalley, D., 14
Smith, L., 298
Starn brothers, 308
Stockhausen, K., 11, 56, 6883, 138, 300,
331, 329
Thomson, J. J., 52
Todoro, T., 114
Truax, B., 97, 112, 115, 190, 3112, 343
Vaggione, H., ix, 3123, 332, 338
Varese, E., 15, 54, 233, 3278, 342
Vetruvius, 52
von Neumann, J., 59
Wagner, R., 11
Weare, C., x
Wiener, N., 63
Wishart, T., 14, 114, 2059, 263, 317
Xenakis, I., ix, 30, 36, 638, 834, 86,
1089, 122, 176, 3003, 329330, 347
Young, T., 50

This page intentionally left blank

Subject Index

Abstract models of particles, 1757

Achorripsis, 83
Acousmagraphe, 163
Acoustic holography, 35
Acoustics, denition of, 6
Aesthetics of microsound, 32548
Agon, 3145
Akron Museum of Art, 308
Aliasing, 32,
Analog electronic instruments, 459
Analogique B, 65, 302
Apple Power Macintosh computer, 111
Arp synthesizer, 329
Art of Noises, 327
As If, 317
Atmospheres, 15
Attracteurs etranges, 318
Attraction and repulsion, 341
Audible Design, 263
Audible intensities, 7
AudioSculpt, 269
Audition, 22
Auditory streams, 23
Basis functions, 2867
Beauty and the Beast, 312
Bell Telephone Laboratories, 244
Bohor I, 15
Buchla synthesizer, 329
Burst masking, 14951

Cantata, 318
Capybara, 114
CCMIX, ix, 383
CDP GrainMill, 114
CEMAMU, 158
Center for Music Experiment. See CME
Center for New Music and Audio
Technologies. See CNMAT
Center for Research in Electronic Art
Technology. See CREATE
Centre de Creation Musicale Iannis
Xenakis. See CCMIX
Centre Nationale de Recherche
Scientique, Marseille. See CNRS
Channel masking, 14951
Chant, 113
Chants de Maldoror, 317
ChaosSynth, 116
Chirplet, 121
Chroma, 132
Clang-Tint, 156, 171, 3079
Cloud Generator, 26, 111, 190, 199200,
224, 319, 383388
Clouds, of sound, 14
bandwidth, 1045
density, 1056
ll factor, 1056
taxonomy, 16
textures, 15
type, 386

404

Subject Index
Clutter, 253
Cmask, 115, 344
CME, 110
CNMAT, 310
CNRS, 284, 288
Coalescence of sound, 277
Code versus grammar, 345
Coherence, 338
Comb wavelets, 289
Comme etrangers et voyageurs sur la terre,
320
Communication versus sensation, 346
Complex tones, 17
Complexity versus simplicity, 343
Composer's Desktop Project, 263
Computer Audio Research Laboratory,
3178
Concerto for Clarinet and Clarinets, 318
Concret PH, 15, 64
Constant-Q granular ltering, 2034
Constraints, 14
Convolution and pulsar synthesis,
2201
Convolution of microsounds, 209221
Convolution of pulsars, 1524
Convolution with clouds, 22431
Cosmophone, 176
CREATE, ix, 193, 224, 232, 321
Creatophone spatializer, 232
Creatovox synthesizer, ix, 1167, 137,
1935, 339, 384
Cri, 320
Cross-ltering, 2167
Csound, 110, 113, 184
Cybernephone, 231
Data packing, 241
Data reduction, 241
De natura sonorum, 15
Decay of information, 348
Dele!, 319
Density of sound, 15, 332
Deterministic plus stochastic analysis,
278
DFT, 2456

Digital Music Systems DMX-1000, 112

Dirac delta function, 367
Discrete Fourier transform. See DFT
Discrete Gabor transform, 296
Disintegration of sound, 277
Dots Points, 157
Dreampaths, 318
Dynamics processing on a micro time
scale, 2045
Dynaphone, 68
Electronic Etude I, 71
Electronic music instruments, historical,
459
Elementa, 318
Eleventh vortex, 156, 3101
EMS Stockholm, 112
En Sueno, 319
Engine 27, 311
Ensoniq sampler, 115
Epitaph fur Aikichi Kuboyama, 61
Eventide H910 Harmonizer, 198
Fast Fourier transform. See FFT
Fast wavelet transform, 284, 287
Feedback loop, 6970
Feuillages, 319
FFT, 239, 244
Field, 3067
Filtering on a micro time scale, 2024
Five Haiku, 323
FM synthesis, 26
FOG, 113, 1635
FOG, 113
Formalism, 336
Formant ltering, 2612
Formant synthesis based on particles,
1638, 177
Fourier analysis, 58
history of, 2434
Fourier series, 244
Fourier transform, 244
Frequency modulation synthesis, 26
Frequency zones, 6
Frequency-domain ltering, 260

405

Subject Index
Gabor analysis, 295249
Gabor matrix, xi, 5763, 65, 92, 296,
349
Gabor transform, 296
musical applications of, 2989
Gaboret, 296
GENDYN, 30, 176
Gesang der Junglinge, 71
GIST, 113
Glisson synthesis, 1215, 177
GMEBaphone, 231
Grain of sound
anatomy, 88
density, 385
duration, 1012, 386
envelope, 8890
generator, 9091
waveform, 1034, 387
Grainlet synthesis, 1259, 177
Grammar versus code, 345
GranQ, 2034
Granula, 114
Granular spatial eects, 107
Granular Synthesis Toolkit, 113
Granular synthesis, 656, 86118, 177
algorithmic models, 978
asynchronous, 92, 967
global organization, 918
implementations of, 108116
parameters of, 99107
physical models, 978
pitch synchronous, 92
quasi-synchronous, 93
spectra of, 989
synchronous, 93
Granulate, 185, 190
Granulation, 187193, 2224
real time, 1912
selective, 191
Graphic synthesis, 15763
GRM, 634, 163
Grossman-Morlet wavelet, 285
Groupe de Recherches Musicale. See
GRM
Gruppen, 75

Half-life, 156, 173, 204, 30910

Harmonizer, 198
Heterogeneity of sound materials, 1819,
335
Hohner Elektronium, 138
Hollerith tabulating machine, 349
Holography, 61
Homogeneity of sound materials, 189
Hymnen, 79
Hyperprism, 262
IBM computer, 313
Idle Chatter, 317
Impulse generators, 129131
Impulse response, 210
Innite time scale, 7
Innitesimal time scale, 36
Infrasound, 7
Instant convolution, 215
Institut Internationale de Musique
Electroacoustique de Bourges, 313
Institute of Sonology, Utrecht, 306, 317
Intensity of sound
audible, 7
perisonic, 7
subsonic, 7
zones of, 6
Intermittency, 3334
International Computer Music
Conference (1980), 306
Intervals versus morphologies, 340
Intuitionism, 336
Invention, 338
Invisible, 318
Ircam Musical Workstation, 113
Ircam, 113, 269
Just_more_idle-chatter, 317
Kitab, 313
Klang-1, 110, 302
Kontakte, 6970, 77, 131
Kronecker delta function, 38
Kunitachi College of Music, 111,
308

406

Subject Index
Kurzweil sampler, 116
Kyma, 114, 322
Laboratoire de Mecanique et
d'Acoustique, 298
La cloche sans vallees, 319
La Maquina de Cantar, 312
Le sacre du printemps, 249
Lemur Pro, 275, 277, 279
Les Ateliers UPIC, 111, 308, 383
Lexicon Varispeech, 198
Licht, 11
Little Animals, 322
Localization blur, 107
Lux Aeterna, 15
Macintosh computer, 111
MacMix, 183
MacPOD, 115
Macro time scale, 11
Macroform, 12
Magic Flute, The, 168
Makbenach, 319
MarcoHack, 27980
Masking, of pulsars, 14951
Massachusetts Institute of Technology.
See MIT
Max, 113
Meso time scale, 14
Metastasis, 64, 83, 123
MetaSynth, 16, 117, 160, 262, 269
Michelson-Stratton harmonic analyzer,
243
Micro arc synthesis, 158, 177
Micro time scale, 20
Micromontage, 1827
Microsound
analog domain, 813
aesthetics of, 325348
modern concept of, 55
perception of, 21
Microtemporal perception
auditory acuity, 24
ssion, 22
fusion, 22

intensity perception, 22
pitch perception, 24
preattentive perception, 24
silence perception, 23
subliminal perception, 25
Mimetismo, 318
MIT, 63, 110
Council for the Arts, 307
Experimental Music Studio, 189, 306
Media Laboratory, 307
Mixed Emotions, 195
MixViews, 318
Modulation with sound particles, 224
Moin-Mor, 321
Moog synthesizer, 313, 329
Morphing, 276
Morphologies versus intervals, 340
Morphology of sound objects, 20
MP3, 242, 322
MPEG 1 layer 3, 242, 322
Multiscale approach to composition,
3301
Multivibrator, 69
Music 4C, 111
Music 11, 110, 189
Music IVBF, 111
Music N languages, 344
Music V, 302
Mycenae-Alpha, 159
NHK Studio, 79
Nodal, 314
Notjustmoreidlechatter, 317
Nscor, 3067
Nyquist frequency, 31
Oberlin Conservatory, 112
Objet sonore, 17
Octuor, 185
Ondes Martenot, 308, 327
Ondioline, 138, 268
Opacity of sound, 322
Oscillator bank resynthesis, 258
Outside time music, 38
Overlap-add resynthesis, 2578

407

Subject Index
Pacic, 312
Pacic Fanfare, 312
Paleo, 318
Parameter variation versus strategy
variation, 343
Particle cloning synthesis, 1713, 177
Particle pluriphony, 231
Particle spatialization, 22133
Particle synthesis, 349
Particle-based formant synthesis, 1638,
177
PAST, 279
Perisonic intensities, 7
Persepolis, 15
Phase vocoder, 239, 2539
Philips Pavilion, 64
PhISEM, 174
Phonogene, 61
Phonogramme, 159161
Phonons, 33
PhotoShop, 269
Photosonic synthesis, 157
Physical modeling synthesis, 26
Physical models of particles, 1735, 177
Pitch-shifting, 200202
Gibson's method, 2012
Lent's method, 2001
Malah, Jones, and Parks's method 200
Pitch-shifting on a micro time scale, 195
6
Pitch-time changing by granulation, 197
202
Pitch-time changing by phase vocoder,
25960
Pitch-time changing by wavelet analysis/
resynthesis, 289
Pithoprakta, 83
Planck time interval, 35
PLFKLANG, 303, 344
Pluriphonic sound, 222
Poeme electronique, 328
Polarons, 33
Presque rien, numero un, 312
Pro Tools, 184, 196
Prototype, 110, 302

Public Organ, 323

Pulsar Studies 120, 323
Pulsar synthesis, 137157, 177, 2201
Pulsaret, 138144
envelope, 140
width modulation, 1424
PulsarGenerator, ix, 138, 154, 323, 339
Pulse width modulation, 142
Punti di tempo, 322
QuickMQ, 277, 279
Real-time convolution, 2156
Real-time granulation, 1912
Recirculating tape feedback loop, 6970
Reection versus spontaneity, 338
Repulsion and attraction, 341
Reverberation by convolution, 2189
Rhythm input by convolution, 21020
Riverrun, 311
Rough versus smooth, 341
Sample time scale, 28
SampleHose, 116
Sawdust, 30
Scambi, 316
Schall, 185, 313
Selective granulation, 191
Semantic network, 346
Sensation versus communication, 346
Short time Fourier transform. See STFT
Sieve theory, 30
Simon Fraser University, 112
Simplicity versus complexity, 343
Size of sounds, 39
SL-9, 321
Smooth versus rough, 341
Soniqworx Artist, 291
Sonographic synthesis, 15763, 177
Sonographic transformation, 263, 26771
Sound, denition of, 6
Sound & Fury, 322
Sound clouds, 14
Sound mass, 14
Sound Mutations, 318

408

Subject Index
Sound object, 17
Sound object morphology, 20
Sound object time scale, 16
Sound Synthesis Program, 29
SoundHack, 204
SoundMagic FX, 116, 190
SoundMaker, 116
Spatialization of particles, 22133
Spatiotemporal eects, 2178
SpecDraw, 269
Spectastics, 116
Spectral mutation, 2612
Spectromorphology, 14
Spontaneity versus reection, 338
SSP, 29
Stable and transient extractor, 2601
Stanford University, 290
Stationary processes, 3334
STFT, 239, 24653
Stochastic masking, 14951
Stochastic processes, 3334
StochGran, 113
Strategy variation versus parameter
variation, 343
Subsample time scale, 31
Subsonic intensities, 7
Sud, 318
SuperCollider, 111, 122, 154
version 2, 115, 154, 190, 344
version 3, 115
Supra time scale, 9
Syd, 175
Synth-O-Matic, 113, 154
Tape feedback loop, 6970
Tar, 313
Telemusik, 79
Telharmonium, 2, 327
Tenth vortex, 156, 3101
Texture-Multiple, 322
The Computer Music Tutorial, viii, 238
The Gates of H., 319
The MIT Press, x
Thema, 185, 187, 313
Thereminovox, 327

They Wash Their Ambassadors in Fennel,

321
Timbre morphing, 276
Time domain, vii, 5
Time point set, 78
Time scales of music, 3
Time-domain harmonic scaling, 200
Time-frequency uncertainty, 250
Tone continuity, illusion of, 330
Tone simultaneity, illusion of, 330
Tones, complex, 17
Tongues of Angels, 312
Tongues of Fire, 317
Touch 'n' go, 322
TPV, 271280
transformations 27480
Tracking phase vocoder. See TPV
Trainlet synthesis, 129137, 177
Traite des objets musicaux, 20
Transformations using wavelets, 28890
Transiciones de Fase, 321
Transient drawing synthesis, 16871, 177
Transparency of sound, 322
TurboSynth, 280
UCSB, ix, 117, 193, 224
UCSD, 109, 305, 3178
Ultrasonic loudspeakers, 33
Ultrasound, 7
Universite de Paris VIII, ix
University of California, Berkeley, 154
University of California, San Diego. See
UCSD
University of California, Santa Barbara.
See UCSB
University of Illinois, UrbanaChampaign, 198
University of Naples Federico II, x,
112
University of York, 113
UPIC, 15860, 320
Variable bit rate, 242
Vector oscillator transform, 2802
Virtual spaces, 222

409

Subject Index
von Neumann lattice, 59
VOSIM generators, 306
VOSIM formant synthesis, 16566
VOX-5, 317
Wavecycle distortion, 2059
Wavelab, 290
Wavelet analysis/resynthesis, 28295
display, 288
ltering, 289
resynthesis, 286
transformations, 28890
Wavelets, 240
Waves versus particles, 44
Waveset distortion, 2059
Window function synthesis, 1678
Window types, 2556
Windowed analysis and transformation,
58, 235300
Windowed frequency domain, vii
Windowed spectrum analysis, 238243
Windowing, 2467
Wings of Nike, 311, 343
Zoetrope, 56
Zones of frequency, 6
Zones of intensity, 6