Audio Watermark Using Matlab

Yiqing Lin · Waleed H.
Abdulla
Audio
Watermark
A Comprehensive Foundation Using
MATLAB
Audio Watermark
Yiqing Lin • Waleed H. Abdulla
Audio Watermark
A Comprehensive Foundation
Using MATLAB
123
Yiqing Lin Waleed H. Abdulla
The University of Auckland The University of Auckland
Auckland, New Zealand Auckland, New Zealand
Additional material to this book can be downloaded from http://extras.springer.com
ISBN 978-3-319-07973-8 ISBN 978-3-319-07974-5 (eBook)

DOI 10.1007/978-3-319-07974-5
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014945572
© Springer International Publishing Switzerland 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Preface
Audio watermarking is a technique providing a promising solution to copyrights

protection for digital audio and multimedia products. Using this technique, hidden
information called watermark containing copyrights information is imperceptibly
embedded into the audio track of a host media. This watermark may be extracted
later on from a suspected media to verify the authenticity. To function as an effective
tool to enforce ownership rights, the audio watermarking scheme must satisfy the
imperceptibility, robustness, security, data payload, and computational complexity
requirements. Throughout this book we will be illustrating in a practical way
the commonly used and novel approaches of audio watermarking for copyrights
protection. We will also introduce our recently developed methods for objectively
predicting the perceptual quality of the watermarked audio signals.
This book is directed towards students, researchers, engineers, multimedia
practitioners, and academics who are interested in multimedia authentication and
audio pirating control. The theoretical descriptions of the watermarking techniques
are augmented by MATLAB implementations to ease understanding of the water-
marking principles. A GUI demonstration program for watermarking embedding
and extraction under different attacks is also provided to quickly surf through the
different aspects of the watermarking attributes.
Book Motivations and Objectives
Motivated by the booming of the digital media applications, plenty of research

has been conducted to investigate the methods of audio watermarking for copy-
rights protection. However, clear and easy to follow information about the audio
watermarking subject are still not widely available and scattered among many
publications. Currently, it is hard to find an easy pathway to develop research in
this field. One main reason to this difficulty is that most of the works are bounded
by IP or patent constraints. On the implementation side it is still hard to find or
write the implementation programs for the known audio watermarking techniques
v
vi Preface
to see how the algorithms work. This book is introduced to establish a shortcut to get
into this interesting field with minimal efforts. The commonly known techniques are
well explained and supplemented with MATLAB codes to get a clear idea about how
each technique performs. In addition, the reader can reproduce the functional figures
of the book with provided MATLAB scripts written specifically for this purpose.
From the robustness and security perspectives, the commonly used audio water-
marking techniques have limitations on the resistance to various attacks (especially
desynchronization attacks) and/or security against unauthorized detection. Thus,
in this book we develop new robust and secure audio watermark algorithm; it is
well explained and implemented in MATLAB environment. This algorithm can
embed unperceivable, robust, blind, and secure watermarks into digital audio files
for the purpose of copyrights protection. In the developed algorithm, additional
requirements such as data payload and computational complexity are also taken
into account and detailed.
Apart from the improvement of audio watermarking algorithms, another land-
mark of this book is the exploration of benchmarking approaches to evaluate
different algorithms in a fair and objective manner. For the application in copyrights
protection, audio watermarking schemes are mainly evaluated in terms of imper-
ceptibility, robustness, and security. In particular, the extent of imperceptibility is
graded by perceptual quality assessment, which mostly involves a laborious process
of subjective judgment. To facilitate the implementation of automatic perceptual
measurement, we explore a new method for reliably predicting the perceptual
quality of the watermarked audio signals. A comprehensive evaluation technique
is illustrated to let the readers know how to pinpoint the strengths and weaknesses
of each technique. The evaluation techniques are supported with tested MATLAB
codes.
Furthermore to what we have just stated that this book extensively illustrates
several commonly used audio watermarking algorithms for copyrights protection
along with the improvement of benchmarking approaches, we may pinpoint the
following new contributions of the current book:
• We introduce a spread spectrum based audio watermarking algorithm for copy-
rights protection, which involves Psychoacoustic Model 1, multiple scrambling,
adaptive synchronization, frequency alignment, and coded-image watermark.
In comparison with other existing audio watermarking schemes [1–10], the
proposed scheme achieves a better compromise between imperceptibility, robust-
ness, and data payload.
• We design a performance evaluation which consists of perceptual quality assess-
ment, robustness test, security analysis, estimations of data payload, and com-
putational complexity. The presented performance evaluation can serve as one
comprehensive benchmarking of audio watermarking algorithms.
• We portray objective quality measures adopted in speech processing for per-
ceptual quality evaluation of audio watermarking. Compared to traditional
perception modelling, objective quality measures provide a faster and more
Preface vii
efficient method of evaluating the watermarked audio signals relative to host

audio signals.
• We analyze methods for implementing psychoacoustic models in the MPEG stan-
dard, with the goal of achieving inaudible watermarks at a lower computational
cost. With the same level of minimum masking threshold, Psychoacoustic Model
1 requires less computation time than Psychoacoustic Model 2.
• We identify the imperceptibility, robustness, and security characteristics of audio
watermarking algorithms and further use them as attacks in the process of
multiple watermarking.
• We propose the use of variable frame length to make the investigated cepstrum
domain watermarking, wavelet domain watermarking, and echo hiding robust
against time-scale modification.
Organization of the Book
The chapters in this book are organized as follows.

Chapter 1 provides an overview of digital watermarking technology and then
opens a discussion on audio watermarking for copyrights protection.
Chapter 2 describes the principles of psychoacoustics, including the anatomy of
the auditory system, perception of sound, and the phenomenon of auditory masking.
Then two psychoacoustic models in the MPEG-1 standard, i.e., Psychoacoustic
Model 1 and 2, are investigated. Through comparisons of the masking effect and the
computational cost, the minimum masking threshold from Psychoacoustic Model 1
is chosen to be used for amplitude shaping of the watermark signal in Chap. 4.
Chapter 3 begins with the implementation specifications for perceptual quality
assessment and the basic robustness test used in this chapter. Then it describes
and evaluates several algorithms for audio watermarking, such as least significant
bit modification, phase coding, spread spectrum watermarking, cepstrum domain
watermarking, wavelet domain watermarking, echo hiding, and histogram-based
watermarking. In the meantime, possible enhancements are exploited to improve
the capabilities of some algorithms.
Chapter 4 presents a spread spectrum based audio watermarking algorithm for
copyrights protection, which uses Psychoacoustic Model 1, multiple scrambling,
adaptive synchronization, frequency alignment, and coded-image watermark. The
basic idea is to embed the watermark by amplitude modulation on the time–
frequency domain of the host audio signal and then detect the watermark by
normalized correlation between the watermarked signal and corresponding secret
keys.
In Chap. 5, the performance of the proposed audio watermarking algorithm
is evaluated in terms of imperceptibility, robustness, security, data payload, and
computational complexity. The evaluation starts with perceptual quality assessment,
which consists of the subjective listening test (including the MUSHRA test and
SDG rating) and the objective evaluation test (including the ODG by PEAQ and
viii Preface
the SNR value). Then, the basic robustness test and the advanced robustness test
(including a test with StirMark for Audio, a test under collusion, and a test under
multiple watermarking) are carried out. In addition, a security analysis is followed
by estimations of data payload and computational complexity. At the end of this
chapter, a comparison between the proposed scheme and other reported systems is
also presented.
Chapter 6 presents an investigation of objective quality measures for perceptual
quality evaluation in the context of different audio watermarking techniques. The
definitions of selected objective quality measures are described. In the experiments,
two types of Pearson correlation analysis are conducted to evaluate the performance
of these measures for predicting the perceptual quality of the watermarked audio
signals.
Auckland, New Zealand Yiqing Lin

Auckland, New Zealand Waleed H. Abdulla
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Information Hiding: Steganography and Watermarking . . . . . . . . . . . . . . 1
1.2 Overview of Digital Watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Framework of the Digital Watermarking System . . . . . . . . . . . . . 4
1.2.2 Classifications of Digital Watermarking . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 Applications of Digital Watermarking . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3.1 Copyrights Protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3.2 Content Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3.3 Broadcast Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3.4 Copy Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Audio Watermarking for Copyrights Protection . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Requirements for the Audio Watermarking System . . . . . . . . . . 8
1.3.1.1 Imperceptibility. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1.2 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1.3 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1.4 Data Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1.5 Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 Benchmarking on Audio Watermarking Techniques . . . . . . . . . . 10
1.3.2.1 Perceptual Quality Assessment . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2.2 Robustness Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.2.3 Security Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Principles of Psychoacoustics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 Physiology of the Auditory System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.1 The Outer Ear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.2 The Middle Ear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.3 The Inner Ear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Sound Perception Concepts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Sound Pressure Level and Loudness . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.2 Hearing Range and Threshold in Quiet . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.3 Critical Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
ix
x Contents
2.3 Auditory Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.3.1 Simultaneous Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1.1 Narrowband Noise Masking Tone . . . . . . . . . . . . . . . . . . . 28
2.3.1.2 Tone Masking Tone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.1.3 Narrowband Noise or Tone Masking
Narrowband Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.2 Nonsimultaneous Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3.2.1 Pre-masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3.2.2 Post-masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Psychoacoustic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.1 Modelling the Effect of Simultaneous Masking . . . . . . . . . . . . . . . 33
2.4.1.1 Models for the Spreading of Masking . . . . . . . . . . . . . . . 33
2.4.1.2 Implementation of Psychoacoustic
Model 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4.1.3 Comparison Between Psychoacoustic
Model 1 and Model 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.2 Modelling the Effect of Nonsimultaneous Masking . . . . . . . . . . 47
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Audio Watermarking Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Specifications on Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1.1 Audio Test Signals Used for Evaluation . . . . . . . . . . . . . . . . . . . . . . . 52
3.1.2 Implementation of Perceptual Quality Assessment . . . . . . . . . . . 53
3.1.3 Implementation of Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.1.3.1 Basic Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.1.3.2 Advanced Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Audio Watermarking Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2.1 Least Significant Bit Modification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2.1.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2.1.2 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.2 Phase Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2.2.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2.3 Spread Spectrum Watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.3.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.4 Cepstrum Domain Watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2.4.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.4.2 Strategies for Improvement . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.5 Wavelet Domain Watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.5.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2.6 Echo Hiding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.6.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Contents xi

3.2.7
Histogram-Based Watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.2.7.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4 Proposed Audio Watermarking Scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.1.1 Selection of Watermarking Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.1.2 Structure of the Watermarking Domain . . . . . . . . . . . . . . . . . . . . . . . 97
4.1.3 Gammatone Auditory Filterbank. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.2 Watermark Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2.1 Embedding Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2.2 Multiple Scrambling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.3 Watermark Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.3.1 Basic Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3.2 Adaptive Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.3 Frequency Alignment Towards Excessive PITSM
and TPPSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.3.3.1 Frequency Alignment Against TSM and PSM . . . . . . 113
4.3.3.2 Implementation of Frequency Alignment . . . . . . . . . . . 114
4.3.3.3 Error Analysis Associated with TBER . . . . . . . . . . . . . . . . 116
4.4 Coded-Image Watermark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5 Performance Evaluation of Audio Watermarking. . . . . . . . . . . . . . . . . . . . . . . . 123
5.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2 Perceptual Quality Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.2.1 Subjective Listening Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.2.2 Objective Evaluation Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.3 Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.3.1 Error Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.3.2 Basic Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.3.3 Advanced Robustness Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.3.3.1 Test with StirMark for Audio . . . . . . . . . . . . . . . . . . . . . . . . 139
5.3.3.2 Test Under Collusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.3.3.3 Test Under Multiple Watermarking. . . . . . . . . . . . . . . . . . 144
5.4 Security Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.5 Data Payload and Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.5.1 Estimation of Data Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.5.2 Estimation of Computational Complexity . . . . . . . . . . . . . . . . . . . . . 153
5.6 Performance Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
xii Contents
6 Perceptual Evaluation Using Objective Quality Measures . . . . . . . . . . . . . 159

6.1 Perceptual Quality Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.2 Objective Quality Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.3 Experiments and Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.3.1 Audio Watermarking Techniques Default Settings. . . . . . . . . . . . 164
6.3.2 Subjective Listening Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.3.3 Objective Evaluation Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.3.4 Performance Evaluation Using Correlation Analysis . . . . . . . . . 169
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
A SDMI Standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
B STEP 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
C StirMark for Audio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
D Critical Bandwidth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
E List of Audio Test Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
F Basic Robustness Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
G Nonuniform Subbands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
List of Figures
Fig. 1.1 A generic digital watermarking system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Fig. 2.1 Structure of the peripheral auditory system . . . . . . . . . . . . . . . . . . . . . . . . . 16
Fig. 2.2 Average pressure levels at auditory canal entrance
versus free-field pressure, at six azimuthal angles
of incidence. Notes: (1) The sound pressure was
measured with a probe tube located at the left ear of
the subject. (2) A point source of sound was moved
around a horizontal circle of radius 1 m with the
subject’s head at the center. At D 0ı , the subject
was facing the source, and at D 90ı , the source was
normally incident at plane of left ear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Fig. 2.3 Anatomy of the cochlea (a) Relative location of the
cochlea in the inner ear (b) Schematic of the unraveled
cochlea (c) Cross-section through one cochlea turn . . . . . . . . . . . . . . . . 20
Fig. 2.4 Resonant properties of the basilar membrane (a)
Envelopes of vibration patterns on the basilar
membrane in response to sound of different
frequencies (b) Distribution of resonant frequencies
along the basilar membrane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Fig. 2.5 Equal-loudness contours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Fig. 2.6 Hearing range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Fig. 2.7 Approximation for the threshold in quiet (a) Frequency
on a linear scale (b) Frequency on a logarithmic scale . . . . . . . . . . . . . 25
Fig. 2.8 Threshold in quiet on Bark scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
xiii
xiv List of Figures
Fig. 2.9 Determination of the critical bandwidth (a) The

threshold for a narrowband noise 2 kHz centered
between two tones of 50 dB as a function of the
frequency separation between two tones (b) The
threshold for a tone of 2 kHz centered between two
narrowband noises of 50 dB as a function of the
frequency separation between the cutoff frequencies of
two noises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Fig. 2.10 Two types of masking: simultaneous and
nonsimultaneous masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Fig. 2.11 Simultaneous masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Fig. 2.12 Masking thresholds for a 60 dB narrowband noise
masker centered at different frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Fig. 2.13 Masking thresholds for a 60 dB narrowband noise
masker centered at different frequencies in Bark scale . . . . . . . . . . . . . 30
Fig. 2.14 Masking thresholds from a 1 kHz narrowband noise
masker at different SPLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Fig. 2.15 Masking thresholds from a 1 kHz narrowband noise
masker at different SPLs in Bark scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Fig. 2.16 Masking thresholds from a 1 kHz tonal masker at
different SPLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Fig. 2.17 Spreading function in ISO/IEC Psychoacoustic Model 1 . . . . . . . . . . 35
Fig. 2.18 Comparison of four spreading functions relative to an
80 dB masker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Fig. 2.19 Initial and normalized PSD estimates (a) Frequency on
linear scale (b) Frequency on Bark scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Fig. 2.20 Tonal and nontonal maskers (a) Frequency on a linear
scale (b) Frequency on Bark scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Fig. 2.21 Individual masking thresholds (a) Frequency on linear
scale (b) Frequency on Bark scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Fig. 2.22 Global masking threshold and minimum masking
threshold (a) Frequency on linear scale (b) Frequency
on Bark scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Fig. 2.23 Mapping between spectral subsamples and subbands . . . . . . . . . . . . . . 46
Fig. 2.24 Comparison of MMTs from Psychoacoustic Model 1
and 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Fig. 2.25 Modelling the effect of post-masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Fig. 3.1 An example of a two-channel stereo signal . . . . . . . . . . . . . . . . . . . . . . . . . 52
Fig. 3.2 Host signal and a watermarked signal by LSB
modification. Note that the watermarked signal is
produced by using L D 6 and modifying the third
and fourth decimal places. (a) Host audio signal. (b)
Watermarked audio signal. (c) Difference between the
watermarked and host audio signals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
List of Figures xv
Fig. 3.3 Host signal and a watermarked signal by the modified

phase-coding method. Note that the watermarked
signal is produced by watermarking with N D 2; 048
and ne D 128. (a) Host audio signal. (b) Watermarked
audio signal. (c) Difference between the watermarked
and host audio signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Fig. 3.4 Block diagram of basic SS watermarking scheme. (a)
Embedding process. (b) Detection process . . . . . . . . . . . . . . . . . . . . . . . . . 63
Fig. 3.5 Host signal and a watermarked signal by SS
watermarking. Note that the watermarked signal is
produced by watermarking with N D 4; 096, nr D 3
and ˇ D 0:03. (a) Host audio signal. (b) Watermarked
audio signal. (c) Difference between the watermarked
and host audio signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Fig. 3.6 Block diagram of computing the complex cepstrum
and the inverse complex cepstrum. (a) Complex
cepstrum xO .n/ D F1 flog . F fx .n/g/g. (b) Inverse
complex cepstrum x .n/ D F1 fexp . F fxO .n/g/g . . . . . . . . . . . . . . . . . . 67
Fig. 3.7 Distributions of Rone and Rzero under different attacks.
Note that these data are produced by watermarking
with N D 4; 096, ˛w D 0:001, and nr D 3 . . . . . . . . . . . . . . . . . . . . . . . . . 70
Fig. 3.8 Host signal and a watermarked signal by cepstrum
domain watermarking. Note that the watermarked
signal is produced by watermarking with N D 2; 048,
˛w D 0:0015 and nr D 3. (a) Host audio signal. (b)
Fig. 3.9 A three-level DWT decomposition and reconstruction.
(a) Wavelet decomposition. (b) Wavelet reconstruction. . . . . . . . . . . . 76
Fig. 3.10 Host signal and a watermarked signal by wavelet
domain watermarking. Note that the watermarked
signal is produced by watermarking with N D 2; 048,
nr D 3, and ˛w D 0:01. (a) Host audio signal. (b)
Fig. 3.11 Impulse response of echo kernels. (a) “One” kernel.
(b) “Zero” kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Fig. 3.12 Auto-cepstrum for echo detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Fig. 3.13 Host signal and a watermarked signal by echo hiding.
Note that the watermarked signal is produced by
watermarking with N D 4; 096, ˛ D 0:2, and nr D 3.
(a) Host audio signal. (b) Watermarked audio signal.
(c) Difference between the watermarked and host audio signals . . 85
xvi List of Figures
Fig. 3.14 Host signal and a watermarked signal by

histogram-based watermarking. Note that the
watermarked signal is produced by watermarking with
Nw D 40, D 2:2, and Eh D 1:5. (a) Host audio
signal. (b) Watermarked audio signal. (c) Difference
between the watermarked and host audio signals . . . . . . . . . . . . . . . . . . . 90
Fig. 4.1 Selection of watermarking regions (a) ET D 0:01 (b) ET D 0:1. . 97
Fig. 4.2 Diagram of blocks in the watermarking domain . . . . . . . . . . . . . . . . . . . . 97
Fig. 4.3 Configuration of one block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Fig. 4.4 Distribution of the watermark bits and synchronization
bit. Note: Slots, , and are used for embedding
B1 , B2 , and Bs , respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Fig. 4.5 Frequency response of a 16-channel GTF . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Fig. 4.6 Block diagram of watermarking one host frame . . . . . . . . . . . . . . . . . . . . 102
Fig. 4.7 Host signal and a watermarked signal by the proposed
scheme. (a) Host signal. (b) Watermark signal. (c)
Watermarked signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Fig. 4.8 Block diagram of watermark detection. Notes: (1)
Basic detection works independently. (2) Adaptive
synchronization is an improvement technique for block
synchronization. (3) Frequency alignment indicated by
dashed lines is an additional solution to excess PITSM
and TPPSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Fig. 4.9 Illustration of random samples cropping . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Fig. 4.10 Flowchart of adaptive synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Fig. 4.11 Duality between TSM and PSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Fig. 4.12 Illustration of frequency alignment (a) Positive PSM
(b) Negative PSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Fig. 4.13 Coded-image denoising by morphological operations . . . . . . . . . . . . . 119
Fig. 4.14 Character recognition by the neural network. (a)
Letters “C” “O” “P” “Y” “R” “I” “G” “H” “T”.
(b) Noisy coded-image watermark. (c) Recovered
coded-image watermark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Fig. 5.1 Determination of watermark strength. ˛w (a) ˛w D 50
for Bass:wav. (b) ˛w D 100 for P op:wav . . . . . . . . . . . . . . . . . . . . . . . . . 126
List of Figures xvii
Fig. 5.2 Screenshot of the MATLAB GUI for the MUSHRA

test. The buttons on the GUI have the following
functions: “Load,” load the host audio signal to be
evaluated. “Start,” start playing the sound from the
beginning. “Pause/Stop,” pause or stop the sound that
is currently playing. “Resume,” resume the sound
from the pause position. “Save,” save the host signal
name and the participant name as well as the registered
scores into a .txt file. “Reset,” reset the interface for
the next trial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Fig. 5.3 Results of the MUSHRA-based subjective listening test . . . . . . . . . . . 130
Fig. 6.1 Objective evaluation via perception modelling . . . . . . . . . . . . . . . . . . . . . 161
Fig. 6.2 Evaluation of PEMO-Q, PEAQ, and EAQUAL . . . . . . . . . . . . . . . . . . . . 167
Fig. 6.3 Overall correlation coefficients over audio test signals
A1 , A2 , and A3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
List of Tables
Table 1.1 Classifications of digital watermarking. . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Table 1.2 Subjective difference grade (SDG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Table 3.1 Results of performance evaluation of LSB modification . . . . . . . . . 59
Table 3.2 Results of performance evaluation of phase coding . . . . . . . . . . . . . . 62
Table 3.3 Results of performance evaluation of SS watermarking . . . . . . . . . . 67
Table 3.4 Results of performance evaluation of cepstrum
domain watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Table 3.5 Results of performance evaluation of wavelet domain
watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Table 3.6 Results of performance evaluation of echo hiding . . . . . . . . . . . . . . . . 86
Table 3.7 Results of performance evaluation of histogram-based
watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Table 5.1 Results of the SDG-based subjective listening test . . . . . . . . . . . . . . . 131
Table 5.2 Results of the objective evaluation test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Table 5.3 Results of error probabilities under TBER D 20 % . . . . . . . . . . . . . . . . 132
Table 5.4 Results of the basic robustness test on the
watermarked Bass.wav signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
watermarked Gspi.wav signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
watermarked Harp.wav signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
watermarked Pop.wav signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Table 5.8 Results of combined attacks on the watermarked
Bass.wav signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Table 5.9 Results of StirMark for Audio attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Table 5.10 Results of the averaging collusion attack . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Table 5.11 Results of multiple self-watermarking. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Table 5.12 Results of inter-watermarking Experiment I . . . . . . . . . . . . . . . . . . . . . . 149
xix
xx List of Tables
Table 5.13 Results of inter-watermarking Experiment II on

Bass.wav signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Table 5.14 Results of the computational complexity estimation . . . . . . . . . . . . . 153
Table 5.15 Performance comparison of different audio
watermarking schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Table 6.1 Comparison of the total computation time (s). . . . . . . . . . . . . . . . . . . . . 168
Table 6.2 Pearson correlation coefficients under our proposed scheme . . . . 171
Table 6.3 Pearson correlation coefficients under cepstrum
Table 6.4 Pearson correlation coefficients under wavelet
Table 6.5 Pearson correlation coefficients under echo hiding . . . . . . . . . . . . . . . 172
Table 6.6 Pearson correlation coefficients under
histogram-based watermarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Table A.1 Robustness test items in SDMI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Table B.1 Robustness test items in STEP 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Table C.1 Robustness test items in StirMark for Audio . . . . . . . . . . . . . . . . . . . . . 181
Table D.1 Critical bands over the frequency spectrum [11]. . . . . . . . . . . . . . . . . . 185
Table E.1 Descriptions of audio test files for performance
evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Table F.1 Descriptions of basic robustness test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Table G.1 Thirty-two nonuniform subbands over the frequency spectrum . . 191
Chapter 1
Introduction
Since the last decade, online distribution of digital multimedia including images,
audio, video, and documents has proliferated rapidly. In the open environment, it is
easy to get free access to various information resources. Along with the convenience
and high fidelity by which digital formatted data can be copied, edited, and
transmitted, massive amounts of copyright infringements have arisen from illegal
reproduction and unauthorized redistribution, which hinders the digital multimedia
industry from progressing steadily [12]. To prevent these violations, the enforcement
of ownership management has become an urgent necessity and is claiming more and
more attention. As a result, digital watermarking has been proposed to identify the
owner or distributor of digital data for the purpose of copyrights protection.
This chapter serves as an overall introduction to the book. First of all, background
knowledge on information hiding, focusing on the differences between steganogra-
phy and watermarking, is presented to ascertain the essence of watermarking. Then
an overview of digital watermarking technology, including system framework, clas-
sifications, and applications, is introduced. Afterward, we focus on the requirements
and benchmarking of audio watermarking for copyrights protection.
1.1 Information Hiding: Steganography and Watermarking
Information hiding is a general concept of hiding data in content. The term “hiding”
can be interpreted as either keeping the existence of the information secret or making
the information imperceptible [13]. Steganography and watermarking are two
important subdisciplines of information hiding. Steganography seeks for ways to
make communication invisible by hiding secrets in a cover message, whereas water-
marking originates from the need for the copyrights protection of the content [14].
The word steganography is derived from the Greek steganos & graphia, which
literally mean “covered writing.” As defined in [13], steganography refers to the
practice of undetectably altering a cover to embed a secret message, i.e., conveying
hidden information in such a manner that nobody apart from the sender and intended
Y. Lin and W.H. Abdulla, Audio Watermark: A Comprehensive Foundation Using MATLAB, 1
DOI 10.1007/978-3-319-07974-5__1, © Springer International Publishing Switzerland 2015
2 1 Introduction
recipient suspects the very existence of the message. Steganography has been used
in a number of ways throughout time, for example, hidden tattoos, invisible inks,
microdots, character arrangement, null ciphers, code words, covert channels, and
spread spectrum communication [15, 16].
Note that steganography appears to be akin to cryptography, but not synonymous.
Both cryptography and steganography are means to provide secrecy, but their
methods of concealment are different. In cryptography, the message is encrypted
to protect its content. One can tell that a message has been encrypted, but cannot
decrypt it without the proper cipher. Once the data are decrypted, the protection is
removed and there is no privacy any longer. In steganography, the message exists,
but its presence is unknown to the receiver and others, such as the adversary. It
is due to this lack of attention that the secret is well preserved. As stated in [15],
“A cryptographic message can be intercepted by an eavesdropper, however, the
eavesdropper may not even know a steganographic message exists.” Therefore,
steganography not only protects confidential information, as does cryptography,
but also keeps the communicating parties safe to some extent. In the meantime,
steganography and cryptography can be combined to provide two levels of security.
That is, we encrypt a message using cryptography and then hide the encryption
within the cover using steganography. This notion can be adopted in digital
watermarking system to increase security.
Watermarking refers to the practice of imperceptibly altering an object to embed
a message about that object [13], i.e., hiding specific information about the object
without noticeable perceptual distortion. Watermarking has a long history dating
back to the late thirteenth century, when “watermarks” were invented by paper
mills in Italy to indicate the paper brand or paper maker and also served as the
basis of authenticating paper. By the eighteenth century, watermarks began to be
used as anticounterfeiting measures on money and other documents. So far, the
most common form of paper watermark remains the bill in many countries. The
first example of a technology similar to our notion of watermarks—imperceptible
information about the objects in which they are embedded—was a patent filed for
“watermarking” musical works by Emil Hembrooke in 1954. He inserted Morse
code to identify the ownership of music, so that any forgery could be discerned. The
term “digital watermarking” is the outcome of the digital era, which appears to have
been first used by Komatsu and Tominaga in 1988. Since 1995, digital watermarking
has gained a lot of attention and has evolved very fast [13, 14].
Watermarking and steganography are two areas of information hiding with differ-
ent emphases. Both of them are required to be robust to protect the secret message.
However, secrecy in watermarking is not strictly necessary, whereas steganography
has to be secret by definition. For instance, it is preferred that everybody knows the
presence of the watermark on bills and can recognize it easily against the light.
Steganography requires only limited robustness as it generally relates to covert
point-to-point communication between trusting parties, while watermarking must
be quite robust to resist any attempts at removing the secret data as it is open
to the public. Furthermore, the concealed message in watermarking is related to
the object which has the same importance as itself. Therefore, no deterioration
1.2 Overview of Digital Watermarking 3
of the perceptual quality of the object is desired. But this is not compulsory in
steganography, because the object there may be merely a carrier and has no intrinsic
value [13, 14].
In the next section, we focus on digital watermarking, that is, watermark-
ing applied to digital data. Key aspects will be discussed towards a deeper
understanding.
1.2 Overview of Digital Watermarking
Digitization over all fields of technology has greatly broadened the notion of
watermarking, and many new possibilities have been opened up. In particular, it
is possible to hide information within digital image, audio, and video files in an
unperceived and statistically undetectable sense. Driven by concerns over digital
rights management (DRM), a new technique called digital watermarking has been
put forward for intellectual property and copyrights protection [17, 18]. Digital
watermarking is not designed to reveal the exact relationship between copyrighted
content and the users, unless one violates its legal use.
Digital watermarking is the process of imperceptibly embedding watermark(s)
into digital media as permanent signs and then extracting the watermark(s) from
the suspected media to assure the authenticity [19]. The watermark(s) is always
associated with the digital media to be protected or to its owner, which means
that each digital media has its individual watermark or each owner has his/her
sole watermark. For the purpose of copyrights protection, the advantage of digital
watermarking over traditional steganography and cryptography is that digital media
can be used in an overt manner, despite the presence of watermarks. In other words,
we do not restrict the access to the watermarks residing in digital media, but make
extra efforts to enhance their robustness against various attacks.
It is worth mentioning that some researchers provide another term closely related
to the issue of copyrights protection, the so-called digital fingerprinting [20–23].
Fingerprints are characteristics of an object that tend to distinguish it from other
similar objects. In a strict sense, fingerprinting refers to the process of identifying
and recording fingerprints that are already intrinsic to the object1 [14]. It is often
regarded as a form of forensic watermarking used to trace authorized users who
distribute them illicitly, i.e., the traitor tracing problem. Note that the greatest
differences between digital watermarking and digital fingerprinting are the origin
of hidden messages and operating mode. In digital watermarking, the watermark
is an arbitrary message containing the information on proprietorship, while the
fingerprint in digital fingerprinting is derived from the host itself and converted
into a unique but much shorter number or string. Essentially, digital fingerprinting
1
Although fingerprinting sometimes is related to the practice of extracting inherent features that
uniquely identify the content, we avoid using this term to prevent confusion [13].
4 1 Introduction
Original Watermark
watermark key
wo kw
Watermark
generator
Watermark ws
signal Attacks AT ( )
Watermarked Attacked Extracted
Host signal signal signal watermark
Watermark Transmission & Watermark
so embedder s communication sa detector we
w
Secret key
ks
Watermark embedding Watermark detection
Fig. 1.1 A generic digital watermarking system
produces a metafile which describes the contents of the source file, so that a piece
of work can be easily found and compared against other works in the database
[17]. For this reason, digital fingerprinting was initially conceived for use in high-
speed searching. Somewhat differently, digital watermarking stemmed from the
motivation for copyrights protection of digital multimedia products. It is able to
stand alone as an effective tool for copyright enforcement.
1.2.1 Framework of the Digital Watermarking System
In general, a digital watermarking system consists of three fundamental parts,

namely a watermark generator, an embedder, and a detector, as illustrated in Fig. 1.1.
Note that for different digital watermarking systems, the inputs indicated by dashed
lines are optional.2
As a rule, the digital media to be protected is called a host signal, so , in which we
choose to embed the original watermark, wo . The form of the original watermark is
diverse; possibly an image, a sequence of letters, or a simple series of bits. wo can be
rearranged into a collection of bits and further processed by a watermark generator
in accordance with the watermark key, kw , so as to generate the watermark signal,
ws . Usually, kw is used as a kind of cryptography to offer additional protection.
Then, the watermark embedder incorporates the watermark signal into the host
2
Hereafter, if the items in the graphs are indicated by dashed lines, it means that they are optional.
Table 1.1 Classifications of digital watermarking

Basis for classification Category
Type of medium to be watermarked Image/ audio/ video/ text
Perceptibility Imperceptible/ perceptible
Robustness of the watermark Robust/ semi-fragile/ fragile
Need of host signal in the detection Blind (public)/ non-blind (private)
Reversibility Nonreversible/ reversible
signal, where the secret key ks is employed to provide extra security and outputs
the watermarked signal sw . The embedding process is mathematically described as
follows:
sw D Embedding .so ; wo ; kw ; ks / (1.1)
where sw should be perceptually similar to so .

After that, the watermarked signal is spread out for communication. During the
course of transmission, sw is likely to be modified in some way, either being pro-
cessed by common signal codings or tampered with by malicious attempts to remove
the watermark. Such modifications are known collectively as “Attacks,” AT ./, for
instance, noise addition, MP3 compression, and random samples cropping.
In the detection, the watermark detector extracts the watermark from the signal
received. The input to the watermark detector is called the attacked signal, sa . The
name is a general term, and sa could be an identical or distorted version of sw . The
detection process is defined by
we D Detection .sa ; so ; kw ; ks / (1.2)
where we is the extracted watermark. By comparing we with wo , it is verified

whether the host signal has been watermarked or not.
1.2.2 Classifications of Digital Watermarking
In terms of different characteristics, digital watermarking can be classified into

several categories as summarized in Table 1.1[24–27].
• Image, audio, video, or text watermarking
There are different kinds of digital media that can be watermarked, such as image,
audio, video, and text document. Image watermarking has developed well since
the beginning of watermarking research. With relation to image watermarking,
most current techniques for video watermarking treat video frames as a sequence
of still images and watermark each of them accordingly. Compared to image
and video watermarking, audio watermarking presents a special challenge due
6 1 Introduction
to less redundancy in audio files and the high sensitivity of the human auditory
system (HAS). With the rapid development of audio compression techniques,
audio products are becoming ever more popular on the Internet. Therefore, audio
watermarking has attracted more and more attention in recent years. Text document
watermarking also has applications wherever copyrighted electronic documents are
distributed [17, 24, 28].
• Imperceptible or perceptible
For images and video, perceptible watermarks are visual patterns such as the logos
merged into one corner of the images, ocular but not obstructive. Although percep-
tible watermarking is easy for practical implementation, it is not the focus of digital
watermarking. As defined before, digital watermarking intends to imperceptibly
embed the watermark into digital media [25].
• Robust, semi-fragile, or fragile
Watermark robustness accounts for the capability of the watermark to survive
various manipulations. A robust watermark is a watermark that is hard to remove
without deterioration of the original digital media. It is usually involved in copy-
rights protection, ownership verification, or other security-oriented applications.
Conversely, a fragile watermark is a watermark that is vulnerable to any mod-
ification, mainly for the purpose of data authentication. In a temperate manner,
a semi-fragile watermark is marginally robust and moderately sensitive to some
attacks [24–26].
• Blind (public) or non-blind (private)
Blind (public) digital watermarking does not require the host signal for watermark
detection. On the contrary, digital watermarking that requires the host signal to
extract the watermark is non-blind (private). Generally, watermark detection is more
robust if the original unwatermarked data are available. However, access to the
original host signal can not be warranted in most real-world scenarios. Therefore,
blind watermarking is more flexible and practical [24, 28].
• Nonreversible or reversible
In reversible watermarking, the watermark can be completely removed from the
watermarked signal, thus allowing it to obtain an exact recovery of the host signal.
However, the price of such reversibility implicates some loss of robustness and secu-
rity. Nonreversible watermarking usually introduces a slight but irreversible degra-
dation in the original signal. Watermark reversibility must only be considered in
applications where complete restoration of the host signal is in great request [24,27].
1.2.3 Applications of Digital Watermarking
Digital watermarking can be used in a wide range of applications. There is no

denying that other techniques might be viable alternatives sometimes. However, the
attributes of digital watermarking make it indispensable for certain purposes.
1.2.3.1 Copyrights Protection
The exploration of digital watermarking was driven by the desire for copyrights
protection. The idea is to embed a watermark with copyright information into the
media. When proprietorial disputes happen, the watermark can be extracted as
reliable proof to make an assertion about the ownership. To this end, the watermark
must be inseparable from the host and robust against various attacks intended
to destroy it. Moreover, the system requires a high level of security to survive
the statistical detection. With these properties, the owner could demonstrate the
presence of watermark to claim the copyright on the disputed media. In addition,
since it is not necessary for the watermark to be very long, the data payload for this
application does not have to be high [29, 30].
1.2.3.2 Content Authentication
In authentication application, the objective is to verify whether the content has been
tampered with or not. Since the watermarks undergo the same transformations as
the host media, it is possible to learn something about the occurrences by looking at
the resulting watermarks. For this purpose, fragile watermarks with a low robustness
are commonly employed. If the content is manipulated in an illegal fashion, fragile
watermarks will be changed to reveal that the content is not authentic [13, 14].
1.2.3.3 Broadcast Monitoring
The target of broadcast monitoring is to collect information about the content being
broadcast. This information is then used as the evidence to verify whether the
content was broadcast as agreed or for some other purposes, such as billing or
statistical analysis for product improvement. In this case, the robustness of the
watermark is not a concern due to a lower risk of distortion. Instead, transparent
or unnoticeable watermarks, i.e., imperceptibility, are more required [13, 31].
8 1 Introduction
1.2.3.4 Copy Control
Most applications of digital watermarking, as discussed so far, have an effect only

after the infringement has happened. In the copy control application, the aim is to
prevent people from making illegal copies of copyrighted media. The mechanism is
to embed watermarks indicating copy status of the content in copyright compliant
devices, proposed by the Secure Digital Music Initiative (SDMI). For example, if the
DVD system contains the data with copyright information embedded as watermarks,
then a compliant DVD player will not play back or copy data that carry a “copy
never” watermark [13, 14, 30].
Digital watermarking also has been applied in device control, legacy enhance-
ment, transaction tracking (or fingerprinting), and so on. More details can be found
in [13, 14, 24, 27, 30].
1.3 Audio Watermarking for Copyrights Protection
Compared to image and video watermarking, inserting watermark(s) into digital

audio files is a more arduous task. Generally, the human auditory system is much
more sensitive than the human visual system (HVS), implying that inaudibility is
much more difficult to achieve than invisibility for images. Moreover, audio signals
are represented by far less samples per time interval, and thereby the amount of
information that can be embedded robustly and inaudibly is much lower than for
visual media [28].
Audio watermarking is a promising solution to copyrights protection for digital
audio and multimedia products. To function as an effective tool to enforce ownership
rights, any eligible audio watermarking scheme must meet a number of requirements
to be described in Sect. 1.3.1. The benchmarking of any audio watermarking
technique is measured against these requirements.
1.3.1 Requirements for the Audio Watermarking System
The audio watermarking system for copyrights protection has to comply with
the following main requirements: excellent imperceptibility for preserving the
perceptual quality of the audio file, strong robustness against various attacks,
and high-level security for preventing unauthorized detection. Data payload and
computational complexity are two additional criteria [30].
1.3 Audio Watermarking for Copyrights Protection 9
1.3.1.1 Imperceptibility
Imperceptibility is a prerequisite to practicality. The process of audio watermarking

is considered to be imperceptible or transparent if no differences between the host
and watermarked signals are perceivable. Otherwise it is perceptible or nontranspar-
ent. To preserve the perceptual quality of the watermarked data, a psychoacoustic
model derived from the auditory masking phenomenon will be relied on to deceive
the human perception of digital audio files [32]. Consequently it appears as if there
is nothing added to the host media.
1.3.1.2 Robustness
Robustness is a measure of reliability and refers to the capability of resisting a

variety of unintentional and intentional attacks. In other words, the watermark
detector should be able to extract the watermark from the attacked watermarked
signal. Examples of attacks on audio watermarking include many kinds of signal
processing and coding, such as noise addition, resampling, requantization, MPEG
(Moving Picture Experts Group) compression, random samples cropping3 , time-
scale modification (TSM), and pitch-scale modification (PSM). The last three
attacks belong to desynchronization attacks, which introduce displacement and
heavily threaten the survival of the watermark.
1.3.1.3 Security
Security is a prerequisite to existence. Since the watermarking algorithms are

likely to be open to the public, we should guarantee that the watermarks cannot
be ascertained even by reversing the embedding process or performing statistical
detection [33,34]. In this case, secret keys (usually pseudorandom sequences) and/or
scrambling operations can be adopted to add randomness into the embedding and
detection processes, so that the digital watermarking system is self-secured.
1.3.1.4 Data Payload
Data payload refers to the amount of bits carried within a unit of time [13]. In digital
audio watermarking, it is defined as the number of bits embedded in a one-second
audio fraction, expressed in bit per second (bit/s or bps). Data payload of the audio
watermarking system varies greatly, depending on the embedding parameters and
3
Random samples cropping includes deliberate removal of the header or footer of a signal.
Therefore, the watermark should be spread throughout the entire audio signal.
10 1 Introduction
the embedding algorithm. Copyrights protection applications do not require a high

data payload, only 2 4 bits/s on average [35].
1.3.1.5 Computational Complexity
From a technological point of view, the computational complexity of a watermark-

ing system involves two principal issues of consideration. One is the speed with
which embedding and detection are performed, and the other is the number of
embedders and detectors [13], where the speed is more of concern for us. The
most intuitive way to estimate the speed is to separately measure embedding and
detection time relative to the duration of the host audio. For a fair comparison, the
measurements should be carried out on platforms with the same computational capa-
bilities. Although a real-time and low-delay system is commonly desired, different
applications require different speeds. For the purpose of copyrights protection,
even a commercial product does not care too much about the embedding time.
Conversely, the customers expect to extract the watermark as quickly as possible.
In practice, no one system can fully satisfy all the requirements and some trade-
offs always exist among criteria. Typically, an audio watermarking system can
operate with either excellent imperceptibility or strong robustness, but not both.
In order to ensure the robustness, we embed the watermark(s) into perceptually
important regions or increase the strength of the watermarking. However, such
strategies are liable to cause perceivable distortion to the host signal, which is
against the property of imperceptibility. Moreover, both of them are in close
connection with data payload. If we embed more bits into an audio signal, the
imperceptibility would become worse and the robustness would be stronger[36].
Similar compromises also occur between imperceptibility, robustness, and security.
1.3.2 Benchmarking on Audio Watermarking Techniques
Along with the advancement of audio watermarking techniques, the necessity

for benchmarking various algorithms effectively and comprehensively becomes
imperative [37, 38]. Since appropriate assessment criteria always depend on the
application, it is impractical and inaccurate to develop a universal benchmark for all
kinds of digital watermarking systems [13]. As discussed above, imperceptibility,
robustness, and security are key principles in designing any audio watermarking
scheme for the application of copyrights protection. Accordingly, performance
evaluations in our research are focused on those three aspects.
Table 1.2 Subjective Difference grade Description of impairments

difference grade (SDG)
0 Imperceptible
1 Perceptible but not annoying
2 Slightly annoying
3 Annoying
4 Very annoying
1.3.2.1 Perceptual Quality Assessment
Similar to evaluating the quality of perceptual codecs in the audio, image, and
video fields [39], perceptual quality assessment on the watermarked audio files
is usually classified into two categories: subjective listening tests by human
acoustic perception and objective evaluation tests by perception modelling or quality
measures. Both of them are indispensable to the perceptual quality evaluation of
audio watermarking.
As perceptual quality is essentially decided by human opinion, subjective
listening tests on audiences from different backgrounds are required in most
applications [39]. In subjective listening tests, the subjects are asked to discern
the watermarked and host audio clips. Two popular modes are the ABX test
[40, 41] and the MUSHRA test (i.e., MUlti Stimuli with Hidden Reference and
Anchors) [42], derived from ITU-R Recommendation BS.1116 [43] and BS.1534
[44]4 , respectively. Moreover, the watermarked signal is graded relative to the host
signal according to a five-grade impairment scale (see Table 1.2) defined in ITU-R
BS.5625 . It is known as the subjective difference grade (SDG), which equals to the
subtraction between subjective ratings given separately to the watermarked and host
signals. Therefore, SDG near 0 means that the watermarked signal is perceptually
undistinguished from the host signal, whereas SDG near 4 represents a seriously
distorted version of the watermarked signal.
However, such audibility tests are not only costly and time-consuming, but also
heavily depend on the subjects and surrounding conditions [46]. Therefore, the
industry desires the use of objective evaluation tests to achieve automatic perceptual
measurement. Currently, the most commonly used objective evaluation is perception
modelling, i.e., assessing the perceptual quality of audio data via a stimulant ear,
such as Evaluation of Audio Quality (EAQUAL) [47], Perceptual Evaluation of
Audio Quality (PEAQ) [48], and Perceptual Model-Quality Assessment (PEMO-Q)
[49]. Moreover, objective quality measures are exploited as an alternative approach
to quantify the dissimilarities caused by audio watermarking. For instance, a widely
used quality measure is the signal-to-noise ratio (SNR), calculated as follows [50]:
4
ITU-R: Radiocommunication Sector of the International Telecommunication Union; BS: Broad-
casting service (sound).
5
ITU-R BS.562 has been replaced by ITU-R BS.1284[45].
12 1 Introduction
P
Œso .n/2
SNR .sw ; so / D 10 log10 P n
(1.3)
n Œsw .n/ so .n/2
P P
where n Œso .n/2 is the power of host signal so and n Œsw .n/ so .n/2 is the
power of noise caused by watermarking.
1.3.2.2 Robustness Test
The goal of the robustness test is to test the ability of a watermarking system
resistant to signal modifications in real applications. In the robustness test, various
attacks are applied to the watermarked signal and produce a number of attacked
signals. Then, watermark detection is performed on each attacked signal to check
whether the embedded watermark survives or not. In particular, the detection rate is
denoted by bit error rate (BER), defined in the following equation:
Number of wrong bits between we and wo

BER D 100% (1.4)
Number of bits of wo
A competent robustness test should comprise an extensive range of possible

attacks. Tens of attacks are employed in some popular audio watermarking evalua-
tion platforms, i.e. SDMI standard, STEP 2000 and StirMark for Audio, which are
described in Appendix A, B and C respectively. In summary, typical signal manipu-
lations on audio watermarking schemes are classified into three categories: common
signal operations (such as noise addition, resampling, requantization, amplitude
scaling, low-pass filtering, echo addition, reverberation, MP3 compression, DA/AD
conversion, and combinations of two or more), desynchronization attacks (such as
random samples cropping, jittering, zeros inserting, time-scale modification and
pitch-scale modification), and advanced attacks (such as collusion and multiple
watermarking6 ). In most cases, a robustness test on an audio watermarking system
includes the first two kinds of attacks, while the last kind is only taken into
consideration for some specific applications. Moreover, desynchronization attacks
are more challenging for most audio watermarking systems. Loss of synchronization
would cause mismatch in positions between watermark embedding and detection,
which is disastrous to watermark retrieval [19].
It is worthy of notice that there is a premise for undertaking a robustness test.
That is, the degree of deterioration by attacks should keep within an acceptable
limit, because it is needless for detection to proceed on a watermarked signal
that is already severely destroyed. Therefore, attack parameters should control the
amplitude of noise added and the extent of stretching or shifting within certain
limits.
6
To be described in section 3.1.3.2, multiple watermarking is to embed several watermarks
sequentially.
1.3.2.3 Security Analysis
Security analysis is performed to evaluate the characteristics of security for audio

watermarking systems. Since security is attributed to the randomness merged
by sequences of pseudorandom numbers (PRN) and/or scrambling operations,
an intuitive method of security analysis is to calculate the number of possible
embedding ways. If there are more possible ways of embedding, it would be difficult
for unauthorized detection to ascertain the embedded watermark. This indicates that
the system has a high level of security.
Note that in the performance evaluation, a variety of audio signals have to be
involved to truly verify the properties of the audio watermarking system. The test set
should be representative of a typical range of audio content [13], such as classical,
rock and folk music, vocal and instrumental music, and so on.
Chapter 2
Principles of Psychoacoustics
Psychoacoustics is the science of sound perception, i.e., investigating the statistical

relationships between acoustic stimuli and hearing sensations [51]. This study
aims to build up the psychoacoustic model, a kind of quantitative model, which
could closely match the hearing mechanism. A good understanding of the sensory
response of the human auditory system (HAS) is essential to the development of
psychoacoustic models for audio watermarking, where the perceptual quality of
processed audio must be preserved to the greatest extent.
In this chapter, the basic structure and function of the auditory system, mainly
the peripheral part, are illustrated for the comprehension of human hearing. Then,
the hearing threshold and auditory masking phenomenon are analyzed to pave
the way for deriving the psychoacoustic models. Finally, Psychoacoustic Model 1
in ISO/MPEG standard is implemented to be utilized in our audio watermarking
scheme later on.
2.1 Physiology of the Auditory System
Hearing is the sense by which sound is perceived [52]. Human hearing is performed
primarily by the auditory system, in which the peripheral part is of more relevance
to our study. The peripheral auditory system (the ear, that portion of the auditory
system not in the brain [53]) includes three components: the outer ear, the middle
ear, and the inner ear, as illustrated in Fig. 2.1.
The whole process of capturing the sound through the ear to create neurological
signals is an intricate and ingenious procedure. First, the sound wave travels through
the auditory canal and causes the eardrum to vibrate. This vibration is transmitted
via the ossicles of the middle ear to the oval window at the cochlea inlet. The
movement of the oval window forces the fluid in the cochlea to flow, which results in
the vibration of the basilar membrane that lies along the spiral cochlea. This motion
causes the hair cells on the basilar membrane to be stimulated and to generate neural
16 2 Principles of Psychoacoustics
Fig. 2.1 Structure of the peripheral auditory system [57]
responses carrying the acoustic information. Then, the neural impulses are sent to
the central auditory system through the auditory nerves to be interpreted by the brain
[54, 55].
2.1.1 The Outer Ear
Sounds communicate the auditory system via the outer ear. The pinna and its deep
center portion, the concha, constitute the externally visible part of the outer ear that
serves focusing the sound waves at the entrance of the auditory canal (or auditory
meatus). Since human pinna has no useful muscles, it is nearly immobile. Therefore,
the head must be reoriented towards the direction of acoustical disturbance for a
better collection and localization of sound. The auditory canal (usually 2–3 cm in
length) is a tunnel through which the sound waves are conducted, and it is closed
2.1 Physiology of the Auditory System 17
with the eardrum (or tympanic membrane).1 The eardrum is stretched tightly across
the inner end of the auditory canal and is pulled slightly inward by structures in the
middle ear [58]. Upon travelling through the auditory canal, sound waves impinge
on the eardrum and cause it to vibrate. Then, these mechanical vibrations which
respond to the pressure fluctuations of acoustic stimuli are passed along to the
middle ear.
The outer ear plays an important role in human hearing. The pinna is of great
relevance to sound localization, since it reflects the arriving sound in ways that
depend on the angle of the source. The resonances occurring in the concha and
auditory canal bring about an increase on sound pressure level (SPL) for frequencies
between 1.5 kHz and 7 kHz. The extent of amplification depends on both the
frequency and angle of the incident wave, as indicated in Fig. 2.2. For example, the
gain is about 10–15 dB in the frequency range from 1.5 kHz to 7 kHz at an azimuthal
angle of 45ı . Moreover, the outer ear protects the eardrum and the middle ear against
extraneous bodies and changes in humidity and temperature [59].
2.1.2 The Middle Ear
The eardrum vibrations are transferred through the middle ear to the inner ear.
The middle ear is an air-filled chamber, bounding by the eardrum laterally and
by the oval window of the cochlea medially. It contains three tiny bones known
as the ossicles: the malleus (or hammer), incus (or anvil), and stapes (or stirrup).
These three ossicles are interconnected sequentially and suspended in the middle
ear cavity by ligaments and muscles. As shown in Fig. 2.1, the malleus is fused to
the eardrum and articulates with the incus; the incus is connected to both the other
bones; the stapes is attached to the incus and its footplate fits into the oval window
of the cochlea. The oval window is a membrane-covered opening which leads from
the middle ear to the vestibule of inner ear.
As an interface between the outer and inner ears, the middle ear has two functions.
One function is to serve as an impedance-matching transformer that ensures an
efficient transmission of sound energy. As we know, the outer and middle ear
cavities are filled with air, while the inner ear is filled with fluid. So the passage of
pressure waves from the outer ear to the inner ear involves a boundary between air
and fluid, two mediums with different acoustic impedance.2 In fact, approximately
99.9 % of sound energy incident on air/fluid boundary is reflected back within the
air medium, so that only 0.1 % of the energy is transmitted to the fluid. It means that
1
In this sense, the auditory canal closed with the eardrum at its proximal end has a configuration
as a resonator.
2
Acoustic impedance is a constant related to the propagation of sound waves in an acoustic
medium. Technically, sound waves encounter much less resistance when travelling in air than in
fluid.
+20
+10
q =0⬚
0.2 0.5 1 2 5 7 10 15
-10
AVERAGE PRESSURE LEVELAT EAR CANAL ENTRANCE VS FREE FIELD PRESSURE (DB)
+20
+10
q =45⬚
0.2 0.5 1 2 5 7 10 15
-10
+20
+10
q =90⬚
0.2 0.5 1 2 5 7 10 15
-10
+10
q =180⬚
0
0.2 0.5 1 2 5 7 10 15
-10
+10
0 q =270⬚
-10
0.2 0.5 1 2 5 7 10 15
-20
+10
q =315⬚
0
-10
0.2 0.5 1 2 5 7 10 15
FREQUENCY (KC/S)
Fig. 2.2 Average pressure levels at auditory canal entrance versus free-field pressure, at six
azimuthal angles of incidence [60]. Notes: (1) The sound pressure was measured with a probe tube
located at the left ear of the subject. (2) A point source of sound was moved around a horizontal
circle of radius 1 m with the subject’s head at the center. At D 0ı , the subject was facing the
source, and at D 90ı , the source was normally incident at plane of left ear
if sound waves were to hit the oval window directly, the energy would undergo a loss
of 30 dB before entering the cochlea. To minimize this reduction, the middle ear has
two features to match up the low impedance at the eardrum with high impedance at
the oval window. The first is related to the relative sizes of the eardrum and the stapes
footplate which clings to the oval window. The effective area of the eardrum is about
55 mm2 and that of the footplate is about 3.2 mm2 ; thereupon they differ in size by
a factor of 17 (55 mm2 /3.2 mm2 = 17). So, if all the force exerted on the eardrum is
transferred to the footplate, then the pressure (force per unit area) at the oval window
is 17 times greater than at the eardrum. The second depends on the lever action of the
ossicular chain that amplifies the force of the incoming auditory signals. The lengths
of the malleus and incus correspond to the distances from the pivot to the applied
and resultant forces, respectively. Measurements indicate that the ossicles as a lever
system increases the force at the eardrum by a factor of 1.3. Consequently, the
combined effect of these actions effectively counteracts the reduction caused by
the impedance mismatch [58]. Another function of the middle ear is to diminish
the transmission of bone-conducted sound to the cochlea by muscle contraction.
If these sounds were sent over to the cochlea, they would appear very loud that may
be harmful to the inner ear [61].
2.1.3 The Inner Ear
The inner ear transduces the vibratory stimulation from the middle ear to neural
impulses which are transmitted to the brain. The vestibular apparatus and the
cochlea are the main parts in the inner ear. The vestibular apparatus is responsible
for the sense of balance. It includes three semicircular canals and the vestibule. The
cochlea is the central processor of the ear, where the organ of corti, the sensory
organ of hearing, is located. The cochlea is a spiral-shaped bony tube structure of
decreasing diameter, which coils up 2 34 times around a middle core containing the
auditory nerve, as shown in Fig. 2.3a.3 The duct is filled with almost incompressible
fluids and is enclosed by the oval window (the opening to the middle ear) and the
round window (a membrane at the rear of the cochlea). When the stapes pushes back
and forth on the oval window, the motion of the oval window causes the fluid to flow
and impels the round window to move reciprocally, which lead to the variations of
fluid pressure in the cochlea. The movements of the oval and round windows are
indicated by the solid and dotted arrows in Fig. 2.3a.
Figure 2.3c shows the cross-section through one cochlea turn. Two membranes,
Reissner’s membrane and the basilar membrane, divide the cochlea along the spiral
direction into three fluid-filled compartments: scala vestibuli, scala media, and scala
tympani. The scala vestibuli and scala tympani are merged through a small opening
called helicotrema at the apex, and they contain the same fluid (the perilymph) with
most of the nervous system. The scala media is segregated from other scalae and
contains a different fluid (the endolymph). On the scala media surface of basilar
membrane (BM) lies the organ of corti. The changes of fluid pressure in the cochlea
will cause the BM to deform, so that the hair cells4 on the organ of corti are
3
Note that the cochlea is a cavity within the skull, not a structure by itself [58]. Hence the unraveled
cochlea in Fig. 2.3b is impossible in practice, only for the sake of illustration.
4
The hair cells including the outer and inner hair cells (OHC and IHC) are auditory receptors on
the organ of corti.
b c
Scala vestibuli
Scala media
Helicotrema
Oval
window
Scala tympani
Basilar
Reissner’s membrane
Round
window membrane
Fig. 2.3 Anatomy of the cochlea (a) Relative location of the cochlea in the inner ear [61]
(b) Schematic of the unraveled cochlea (c) Cross-section through one cochlea turn [65]
stimulated to transduce the movement of the BM into neural impulses. Then the
neural signals are carried over to the brain via auditory nerve, which ultimately lead
to the perception of sound.
The basilar membrane extends along the spirals of the cochlea and is about 32 mm
long. It is relatively narrower and stiffer at the base (near the windows), while
it gets wider and more flexible at the apex (near the helicotrema). Accordingly,
each location on the BM has different vibratory amplitude in response to sound of
different frequencies, which means that each point resonates at a specific charac-
teristic frequency (CF) [54]. As exemplified in Fig. 2.4a, for high-frequency tones,
the maximum displacement of the BM occurs near the base, with tiny movement
on the remainder of the membrane. For low-frequency tones, the vibration travels
all the way along the BM, reaching its maximum close to the apex.5 Figure 2.4b
5
There is one fact worth of attention, i.e., any location on the BM will respond to a wide range of
tones that are lower than its CF. That’s why low frequencies are less selective than high frequencies.
Fig. 2.4 Resonant properties of the basilar membrane (a) Envelopes of vibration patterns on the
basilar membrane in response to sound of different frequencies [66] (b) Distribution of resonant
frequencies along the basilar membrane [64]
summarizes the distribution of frequencies that produce maximum displacement at

different positions along the basilar membrane. Note that the spacing of resonant
frequencies is not linear to the frequency, but in a logarithmic scale approximately.
It is called Bark scale or critical band rate corresponding to the concept of critical
bands.
In this sense, the cochlea performs a transformation that maps sound frequencies
onto certain locations along the basilar membrane, i.e., a “frequency-to-place”
conversion [51]. It is of great importance to the comprehension of auditory masking.
Since one frequency maximally excite only one particular point on the basilar
membrane, the auditory system acts as a frequency analyzer which can distinguish
the frequencies from each other. If two tones are different enough in frequency, the
response of the BM to their combination is simply the addition of two individual
ones. That is, there are two vibration peaks along the BM, at the positions identical
to where they would be if two tones were presented independently. However, if two
tones are quite close in frequency, the basilar membrane would fail to separate the
combination into two components, which results in the response with one fairly
broad peak in displacement instead of two single peaks [58]. As for the interval
how far two tones can be discriminated, it depends on critical bands and critical
bandwidths discussed next.
2.2 Sound Perception Concepts
Sounds are rapid variations in pressure, which are propagated through the air away
from acoustic stimulus. Our sense of hearing allows us to perceive sound waves
of frequencies between about 20 Hz and 20 kHz. As discussed in the mechanism
of human ear, perception of sound involves a complex chains of events to read
the information from sound sources. Naturally, we are often surrounded with a
mixture of various sounds and the perception of one sound is likely to be obscured
by the presence of others. This phenomenon is called auditory masking, which is
the fundamental of psychoacoustic modelling. Here, some basic terms related to
auditory masking are introduced.
2.2.1 Sound Pressure Level and Loudness
Sound reaches human ear in the form of pressure waves varying in time, s .t /.
Physically, the pressure p is defined as force per unit area, and the unit in MKS
system is Pascal (Pa) where 1 Pa D 1 N=m2 . Also, the intensity is defined as power
per unit area and its unit is W=m2 . In psychoacoustics, values of sound pressure vary
from 105 Pa (ATH, absolute threshold of hearing) to 102 Pa (threshold of pain). To
cover such a broad range, (SPL) is defined in logarithm units (dB) as
2
p I
LSPL =dB D 10 log10 D 10 log10 ; (2.1)
p0 I0
where LSPL is the SPL of a stimulus, p is the pressure of stimulus in Pa, p0 D

20 Pa is the reference pressure of a tone with frequency around 2 kHz, I is
sound intensity of the stimulus, and I0 D 1012 W=m2 is the reference’s intensity
correspondingly [11].
The hearing sensation that relates to SPL is loudness of sound, expressed in
phon. Note that loudness is a psychological, not a physical, attribute of sound. By
definition, the loudness level of a 1 kHz tone is equal to its SPL in dB SPL [61].
The perceived loudness of sound depends upon its frequency as well as its intensity,
as described by a series of equal-loudness contour in Fig. 2.5. Each equal-loudness
2.2 Sound Perception Concepts 23
FEELING
120 phons
120
110
100
SOUND PRESSURELEVEL (DB)
100
90
80
80
70
60
60
50
40
40
30
20
20
Minimum audible field 10
(MAF)
0
0
20 100 500 1000 5000 10000

FREQUENCY (HZ)
Fig. 2.5 Equal-loudness contours [69]
contour represents SPLs required at different frequencies in order that all tones on
the contour are perceived equally loud [68]. The loudness of 20 phons contour at
100 Hz with 50 dB SPL is perceived similar to 1 kHz with 20 dB SPL. In Fig. 2.5,
the deviation from the maximum sensitivity region of equal-loudness contours at
high phons (i.e., 120 phons) is lower than those of low phons (i.e., 10 phons). This
indicates that the sensitivity to frequency changing of HAS at low phons is relatively
higher than high phons. Hence, complex sounds with identical frequency and phase
components might sound different due to variations in loudness [58].
2.2.2 Hearing Range and Threshold in Quiet
Human hearing spreads widely from 20 Hz to 20 kHz in frequency, as well as

ranging from about 0 dB up to 120 dB in SPL. The most sensitive part is between
100 Hz and 8 kHz for human speech. Figure 2.6 shows hearing range of human,
where different hearing thresholds are sketched in SPL curves as function of
frequency.
The hearing threshold at the bottom is the threshold in quiet, or (ATH), which
approximately corresponds to the baseline in Fig. 2.5. It decreases gradually from
20 Hz to 3 kHz and then increases sharply above 16 kHz. The threshold in quiet
indicates, as a function of frequency, the minimum SPL of a pure tone to be audible
in a noiseless environment. Thus under no circumstances the human ear can perceive
140 140 100 200

W Pa
threshold of pain m2
120 120 1 20
SOUND PRESSURE LEVEL (DB)
lim
100 it o 100 10-2 2
f dam
age risk
SOUND INTENSITY
SOUND PRESSURE
80 music 80 10-4 0.2
60 speech 60 10-6 2.10-2
40 40 10-8 2.10-3
20 threshold 20 10-10 2.10-4

in quiet
0 0 10-12 2.10-5
0.02 0.05 0.1 0.2 0.5 1 2 5 10 20

FREQUENCY (KHZ)
Fig. 2.6 Hearing range [11]
sounds at SPLs below that threshold. In other words, frequency components that
fall below the threshold in quiet are insignificant to our perception of sound and
unnecessary to be processed [51]. This property is crucial to the development
of psychoacoustic model, where the threshold in quiet is approximated by the
following frequency-dependent function:
Threshold in Quiet .f / =dB D
0:8 ( 2 ) 4
f f 3 f
3:64 6:5 exp 0:6 3:3 C 10 (2.2)
1000 1000 1000
as plotted on both linear and logarithmic scales in Fig. 2.7. Regarding Eq. (2.2), one
point to note is that it only applies to the frequency range 20 Hz f 20 kHz.
2.2.3 Critical Bandwidth
As discussed in Sect. 2.1.3, the cochlea performs a “frequency-to-place” conver-

sion and each position on the basilar membrane responds to a limited range
of frequencies. Accordingly, the peripheral auditory system acts as a spectrum
analyzer, modelling as a bank of band-pass filters with overlapping passbands
[61]. Empirically, the main hearing range between 20 Hz and 16 kHz is divided
into 24 nonoverlapping critical bands, and the critical bandwidths (CB) are listed
2.2 Sound Perception Concepts 25
a b
120 120
100 100
Sound pressure level (dB)

80 80
60 60
40 40
20 20
0 0
0.021 2 3.3 5 10 20 0.02 0.1 1 2 3.3 5 10 20

Frequency on a linear scale (kHz) Frequency on a base 10 logarithmic scale (kHz)
Fig. 2.7 Approximation for the threshold in quiet (a) Frequency on a linear scale (b) Frequency
on a logarithmic scale
in Appendix D. We call it critical band rate scale and its unit is Bark. One
Bark represents one critical band and corresponds to a distance along the basilar
membrane of about 1.3 mm.6 Considering nonlinear spacing of resonant frequencies
on the basilar membrane, it is expected that critical bandwidths are nonuniform,
varying as a function of frequency. The following equation describes the dependence
of Bark scale on frequency [11]:
2
0:76fl fl
z=Bark D 13 arctan C 3:5 arctan ; (2.3)
1000 7500
where fl is the lower frequency limit of critical bandwidth. For example, the
threshold in quiet in Fig. 2.7 is plotted on Bark scale as shown in Fig. 2.8.
Note that each critical bandwidth only depends on the center frequency of the
passband. It is demonstrated in Fig. 2.9, where the critical bandwidth at 2 kHz is
measured. As shown in Fig. 2.9a, hearing threshold is flat about 33 dB until two
tones are about 300 Hz away from each other, and then it drops off rapidly. A
similar result is obtained from Fig. 2.9b, hearing threshold is rather flat about 46 dB
until two noises are away from 300 Hz [51]. Consequently, the critical bandwidth is
300 Hz for a center frequency of 2 kHz. It is worth mentioning that the threshold in
Fig. 2.9b is at 46 dB versus only 33 dB in a, which means narrowband noises reduce
6
The whole length of 32 mm basilar membrane divided by 24 critical bands is 1.3 mm for each
band.
120
100
Sound pressure level (dB) 80
60
40
20
0 5 10 1516.2 20 25
Frequency on critical band rate scale (Bark)
Fig. 2.8 Threshold in quiet on Bark scale
a b
Level of narrow-band noise (dB)
50 50
Level of test tone (dB)
Δf 50dB
40 40
f
2kHz
30 30
Δf 50dB
20 20
f
2kHz
10 10
50 100 200 500 1000 2000 50 100 200 500 1000 2000
Difference, Δf, between cut-off frequencies (Hz) Frequency separation of the two tones (Hz)
Fig. 2.9 Determination of the critical bandwidth [11] (a) The threshold for a narrowband noise
2 kHz centered between two tones of 50 dB as a function of the frequency separation between two
tones (b) The threshold for a tone of 2 kHz centered between two narrowband noises of 50 dB as a
function of the frequency separation between the cutoff frequencies of two noises
more audibility than tones. This fact is referred to “asymmetry of masking” and
more details will be discussed in the next section.
On the basis of experimental data, an analytic expression is derived to better
describe critical bandwidth 4f as a function of center frequency fc [11]:
" 2 #0:69
fc
4f =Hz D 25 C 75 1 C 1:4 : (2.4)
1000
2.3 Auditory Masking 27
The concept of critical bandwidth contributes to the understanding of auditory

masking, because CB around a masker denotes the frequency range over which the
main masking effect operates. As demonstrated in Sect. 2.3.1, the masking curves
distribute equally across the spectrum in Bark scale.
2.3 Auditory Masking
Due to the effect of auditory masking, the perception of one sound is related to not
only its own frequency and intensity, but also its neighbor components. Auditory
masking refers to the phenomenon that one faint but audible sound (the maskee)
becomes inaudible in the presence of another louder audible sound (the masker).
It has a great influence on hearing sensation and involves two types of masking,
i.e., simultaneous masking and nonsimultaneous masking (including pre-masking
and post-masking) as displayed in Fig. 2.10. Due to auditory masking, any signals
below these curves cannot be heard. Therefore, by virtue of auditory masking, we
can modify audio signals in a certain way without perceiving deterioration, as long
as the modifications could be properly “masked.” This notion is the essence of audio
watermarking [70, 71].
2.3.1 Simultaneous Masking
Simultaneous masking (or frequency masking) refers to masking between two

sounds with close frequencies, where the low-level maskee is made inaudible by
simultaneously occurring louder masker. Both masker and maskee can be sinusoidal
Pre-masking Simultaneous Post-masking
Masker on
-20 0 100 250 Time (ms)
Fig. 2.10 Two types of masking: simultaneous and nonsimultaneous masking

S0 Masker
Masking
Partially masked
S1 threshold
Threshold sound
in quiet S3
Masked S
2
sounds
0.5 1 4 Frequency (kHz)
Fig. 2.11 Simultaneous masking
tone or narrowband noise.7 Figure 2.11 gives an example of simultaneous masking,

where sound S0 is the masker. Because of the presence of S0 , the threshold in quiet
is elevated to produce a new hearing threshold named as masking threshold. The
masking threshold is a kind of limit for just noticeable distortion (JND) [72], which
means that any sounds or frequency components below this threshold are masked by
the presence of the masker. For instance, the weaker signal S1 and S2 are completely
inaudible, as their SPLs are below the masking threshold. For the signal S3 , it is
partially masked and only the portion above the threshold is perceivable. Moreover,
the effective masking ranges for the maskers at different frequencies are determined
solely by critical bandwidths, as implied in Fig. 2.9. If the maskee lies in critical
band of the masker, the maskee is more likely to be unperceived. The mechanism
by which masking occurs is still uncertain [61]. In general, it is because the louder
masker creates an excitation of sufficient strength on the basilar membrane. Then
such an excitation prevents the detection of another excitation within the same
critical band from a weaker sound [51].
The masking threshold depends on the characteristics of both masker and maskee.
Considering two possibilities of each, there are four cases in simultaneous masking,
that is, narrowband noise masking tone (NMT), tone masking tone (TMT), narrow-
band noise masking narrowband noise (NMN), and tone masking narrowband noise
(TMN).
2.3.1.1 Narrowband Noise Masking Tone
Most often, the case of NMTs happens, where the masker is narrowband noise
and the maskees are tones located in the same critical band. Figure 2.12 shows the
masking thresholds for narrowband noise masker masking tones, where the noise is
7
Here, narrowband means the bandwidth equal to or smaller than a critical band.
80
Level of test tone (dB) 60

fc = 0.25 4kHz
2 dB 3 dB 5 dB
40
20
0.02 0.05 0.1 0.2 0.5 1 2 5 10 20

Frequency of test tone (kHz)
Fig. 2.12 Masking thresholds for a 60 dB narrowband noise masker centered at different frequen-
cies [51]
at a SPL of 60 dB and centered at 0.25, 1, and 4 kHz separately. In the graph, solid
lines represent masking thresholds, and the dashed line at the bottom is the threshold
in quiet.8 The masking thresholds have a number of important features. For example,
the form of curve varies with different maskers, but always reaches a maximum near
the masker’s center frequency. It means that the amount of masking is greatest when
the maskee is located at the same frequency with the masker. The masking ability of
a masker is indicated by the minimum signal to mask ratio (SMR), i.e., the minimum
difference of SPL between the masker and its masking threshold. Therefore, higher
SMR implies less masking. Another point is that low-frequency masker produces a
broader masking threshold and provides more masking than high frequencies. Here,
the 0.25, 1, and 4 kHz thresholds have a SMR of 2, 3, and 5 dB, respectively.
Figure 2.12 is sketched in normal frequency units, where the masking thresholds
of different frequencies are dissimilar in shape. If graphed in Bark scale, all the
masking thresholds look similar in shape as shown in Fig. 2.13.9 In this case, it
is easier to model the masking threshold by the use of the so-called spreading
function in Sect. 2.4.1.1. As a result, Bark scale is widely used in the area of auditory
masking.
Moreover, the masking thresholds from a 1 kHz narrowband noise masker at
different SPLs, LCB , are outlined in Fig. 2.14. Although SPL of the masker is
different, the minimum SMR remains constant at around 3 dB, corresponding to the
value in Fig. 2.12. It means that the minimum SMR in NMT solely depends on the
center frequency of masker. Also notice that the masking threshold becomes more
asymmetric around the center frequency as the SPL increases. At frequencies lower
than 1 kHz, all the curves have a steep rise. But at frequencies higher than 1 kHz,
8
Hereafter, this rule does apply to all the graphs in Sect. 2.3.
9
For illustration, all the curves are shifted upward to the masker’s SPL (60 dB).
Threshold, excitation level (dB)

0.07 0.25 0.5 1 fm = 2kHz 4 8
60
40
20
0 2 4 6 8 10 12 14 16 18 20 22 24
Critical-band rate (Bark)
Fig. 2.13 Masking thresholds for a 60 dB narrowband noise masker centered at different frequen-
cies in Bark scale [51]
100
LCB =
80
Level of test tone (dB)
100 dB
60
80
40
60
20
40
20
0
0.02 0.05 0.1 0.2 0.5 1 2 5 10 20

Fig. 2.14 Masking thresholds from a 1 kHz narrowband noise masker at different SPLs [51]
the slopes of maskers at higher SPLs decrease more gradually. Recall Fig. 2.4a; it is
reasonable to expect that the masker is good at masking the tones whose frequencies
are lower than its own frequency, rather than higher frequency tones [58]. To show
the similarity in shape over all the masking thresholds, Fig. 2.15 plots the curves in
Bark scale again.
2.3.1.2 Tone Masking Tone
The early work on auditory masking started from experiments on tones masking
tones within the same critical band. Since both the masker and maskee are pure
tones, their interference is likely to result in the occurrence of beats. Therefore,
100
Threshold, excitation level (dB)
LG = 100dB
80
80
60
60
40
40
20
20
0
0 2 4 6 8 10 12 14 16 18 20 22 24
Critical-band rate (Bark)
Fig. 2.15 Masking thresholds from a 1 kHz narrowband noise masker at different SPLs in Bark
scale [51]
besides the masker and maskee, additional beating tones become audible and
accordingly disturb the subjects. Figure 2.16 shows the masking thresholds from
a 1 kHz tonal masker at different SPLs. During the course of approaching 1 kHz,
the maskee was set 90ı out of phase with the masker to prevent beating. Similar
to Fig. 2.14, the masking thresholds spread also broader towards high frequencies
than lower frequencies. However, an obvious difference lies in the minimum SMR,
roughly 15 dB in Fig. 2.16 versus about 3 dB in Fig. 2.14. It indicates that the
narrowband noise is a better masker than pure tone, referred as “asymmetry of
masking” [73]. This fact actually has been demonstrated in Fig. 2.9 already. The
masking threshold by narrowband noise masker in Fig. 2.9b is valued at 46 dB,
higher than 33 dB by tonal masker in Fig. 2.9a. So in psychoacoustic modelling,
we should identify the frequency components to be noise-like or tone -like and then
calculate their masking thresholds separately.
2.3.1.3 Narrowband Noise or Tone Masking Narrowband Noise
In contrast to NMT and TMT, it is more difficult to characterize narrowband noise

or tone masking narrowband noise. So far, relatively few studies in NMN and TMN
are carried out. Under the case of NMN, the masking thresholds heavily rely on
phase relationship between the masker and maskee. In other words, different relative
phases between the masker and maskee would lead to different values of minimum
SMRs. It is reported that measurements for wideband noise have minimum SMRs
of about 26 dB [51,73]. As for TMN, the minimum SMR tends to fluctuate between
20 and 30 dB [51].
80
Level of test tone (dB) 60 LM=90dB
40
70
20
50
30
0
0.02 0.05 0.1 0.2 0.5 1 2 5 10 20

Fig. 2.16 Masking thresholds from a 1 kHz tonal masker at different SPLs [51]
2.3.2 Nonsimultaneous Masking
In addition to simultaneous masking, auditory masking can also take place when
the maskee is present immediately preceding or following the masker. This is called
nonsimultaneous masking or temporal masking. As exemplified in Fig. 2.10, one
200 ms masker masks a tone burst with very short duration relative to the masker.
There are two kinds of nonsimultaneous masking: (1) pre-masking or backward
masking, occurring just before the onset of masker, and (2) post-masking or forward
masking, occurring after the removal of masker. In general, the physiological
basis of nonsimultaneous masking is that the auditory system requires a certain
integration time to build the perception of sound, where louder sounds require longer
integration intervals than softer ones [51].
2.3.2.1 Pre-masking
Pre-masking is somewhat unexpected since it happens before the presence of

masker. As seen from Fig. 2.10, the duration of pre-masking is quite short (about
20 ms), whereas it is most effective only in 1–2 ms before the onset of masker [73].
It is suggested that the duration of masker might affect the time that pre-masking
lasts. Up to now, however, no experimental results could specify such a relation.
Pre-masking has less masking capacity than post-masking and simultaneous
masking; nevertheless, it plays a significant role in the compensation of pre-noise or
pre-echo distortion [51].
2.4 Psychoacoustic Model 33
2.3.2.2 Post-masking
Post-masking is better understood compared to pre-masking. It reflects a moderate

decrease of the masking level after the masker is halted. As displayed in Fig. 2.10,
post-masking level decays gradually to zero after a longer period of time (about
150 ms). Therefore, post-masking exhibits a higher masking capacity which is
beneficial to most applications. Experimental studies have revealed that post-
masking depends on the intensity and duration of the masker as well as relative
frequency of the masker and maskee [51].
2.4 Psychoacoustic Model
The knowledge of auditory masking provides the foundation for developing psy-
choacoustic models. In psychoacoustic modelling, we use empirically determined
masking models to analyze which frequency components contribute more to the
masking threshold and how much “noise” can be mixed in without being perceived.
This notion is applicable to audio watermarking, of which the imperceptibility is
one prerequisite. Typically, in some audio watermarking techniques such as spread
spectrum watermarking [74, 75] and wavelet domain watermarking [7, 76], the
watermark signal is added to the host signal as a faint additive noise. To keep
the watermarks inaudible, we often utilize the minimum masking threshold (MMT)
calculated from psychoacoustic model to shape the amplitude of watermark signal.
2.4.1 Modelling the Effect of Simultaneous Masking
Modelling the effect of simultaneous masking is one major task of psychoacoustic

model. In general, there are a series of steps involved. Firstly, the input audio signal
is analyzed to classify its noise-like and tone-like frequency components, due to
the phenomenon of “asymmetry of masking.” Secondly, the so-called spreading
functions are derived to mimic the excitation patterns of noise-like and tone-like
maskers, respectively. Thirdly, after shifted down by a certain amount for each
masker, all the individual masking thresholds as well as ATH are added up in
some manner to obtain a global masking threshold, an estimation on the concurrent
masking effect. Finally, we take the lowest level of global masking threshold in each
frequency band to obtain the (MMT), which represents the most sensitive limit.
2.4.1.1 Models for the Spreading of Masking
Models for the spreading of masking are developed to delineate excitation patterns
of the maskers. As noticed from two examples of excitation patterns in Figs. 2.13
and 2.15, the shape of curves are quite similar and also easy to describe in
Bark scale, because Bark scale is linearly related to basilar membrane distances.
Accordingly, we define spreading function SF.d z/ as a function of the difference
between the maskee and masker frequencies in Bark scale, d z=Bark D z .fmaskee /
z .fmasker /. Apparently, d z 0 when the masker is located at a lower frequency than
the maskee, and d z < 0 when the masker is located at a higher frequency than the
maskee.
There are a number of spreading functions introduced to imitate the characteris-
tics of maskers. For instance, two-slope spread function is the simplest one that uses
a triangular function:
(
Œ27 C 0:37 max fLM 40; 0g d z; d z 0
10 log10 SF .d z/ =dB D (2.5)
27d z; d z < 0;
where LM is SPL of the masker.

Another popular spreading function is proposed by Schroeder and expressed as
the following analytical function:
q
10 log10 SF .d z/ =dB D 15:81 C 7:5 .d z C 0:474/ 17:5 1 C .d z C 0:474/2 :
(2.6)
After slight modification on Schroeder’s spreading function, spreading function as
Eq. (2.7) is adopted in ISO/IEC MPEG10 Psychoacoustic Model 2.
10 log10 SF .d z/ =dB
q
D 15:8111389 C 7:5 .1:05d z C 0:474/ 17:5 1 C .1:05d z C 0:474/2
h i
C8 min 0; .1:05d z 0:5/2 2 .1:05d z 0:5/ : (2.7)
It should be noted that the two spreading functions Eqs. (2.6) and (2.7) are indepen-
dent of the masker’s SPL, which is advantageous to reduction in computation when
generating overall masking threshold.
The spreading function utilized in ISO/IEC MPEG Psychoacoustic Model 1 is
different from Psychoacoustic Model 2:
8
ˆ
ˆ 17d z 0:4LM C 11; 3 d z < 1
ˆ
ˆ
<.0:4L C 6/ d z; 1 d z < 0
M
10 log10 SF .d z/ =dB D : (2.8)
ˆ
ˆ 17d z; 0 dz < 1
ˆ
:̂
17d z C 0:15LM .d z 1/ ; 1 dz < 8
10
ISO: International Organization for Standardization; IEC: International Electrotechnical Com-
mittee; MPEG: Moving Picture Experts Group.
LM=20
100
LM=40
LM=60
LM=80
80
LM=100
60
40
20
0
−4 0 4 8 12
Critical band rate (Bark)
Fig. 2.17 Spreading function in ISO/IEC Psychoacoustic Model 1
Figure 2.17 shows spreading functions in Model 1 for different levels of the masker.
It is seen that the higher SPL the masker has, the more asymmetric the curve looks.
Specifically, higher frequencies exhibit more masking than lower frequencies when
the level of masker is high. This two-piece linear spreading function is a good
approximation to the masking thresholds of TMT in Fig. 2.16.
In addition, four models described above for spreading functions, i.e., two-slope
SF, Schroeder SF, Psychoacoustic Model 1 SF, and Model 2 SF, are compared
at a level of 80 dB in Fig. 2.18. Among these four models, two-slope spreading
function is the most conservative one, and Model 1 spreading function allows for
more upward spreading of masking than others [51].
2.4.1.2 Implementation of Psychoacoustic Model 1
In different application scenarios, psychoacoustic model can be implemented in

different ways to satisfy the criteria required. ISO/IEC MPEG-1 Standard [77]
utilizes two informative psychoacoustic models, Psychoacoustic Model 1 and 2, to
determine the MMT for inaudibility. Typically, Model 1 is applied to MPEG Layers
I and II and Model 2 to MPEG Layer III. Both models are commonly in use and
well performed. Psychoacoustic Model 1 proposed a low-complication method to
analyze spectral data and output SMR, whereas Psychoacoustic Model 2 performs
a more detailed analysis at the expense of greater computational complexity
Two−slope SF
80 Schroeder SF
Model 1 SF
Model 2 SF
60
40
20
0
−8 −4 0 4 8 12
Fig. 2.18 Comparison of four spreading functions relative to an 80 dB masker
[31, 78, 79]. Hence, Psychoacoustic Model 1 for Layer I is later employed in our
audio watermarking scheme in consideration of its higher efficiency.
In our case, the input to Psychoacoustic Model 1 is one frame of audio signal
and the corresponding output is its MMT. The whole procedure of implementation
consists of six steps [72, 73, 77, 80]:
1. FFT analysis and SPL normalization
2. Identification of tonal and nontonal maskers
3. Decimation of invalid tonal and nontonal maskers
4. Calculation of individual masking thresholds
5. Calculation of global masking threshold
6. Determination of the MMT
The details of each step are expounded as follows:
• STEP 1: FFT analysis and SPL normalization

For an accurate analysis of frequency components, fast Fourier transform (FFT) is
performed to obtain a high-resolution spectral estimate of incoming frame x .n/.
In Psychoacoustic Model 1, the input frame has a size of N D 512 points. To
minimize the leakage effect, x .n/ is multiplied with a modified Hanning window
w .n/ defined by
r r
8 8 1 2 n
w .n/ D hann .N / D 1 cos 0 n N 1; (2.9)
3 3 2 N

where hann .N / D 12 1 cos 2Nn is the N -point Hanning window. Factor
q
8
3
is a gain to compensate the average power of w .n/, so that hw .n/2 i
X
N 1 h i
2
1
N
w .n/ D 1. Then, power spectral density (PSD) of x .n/ is computed as
nD0
ˇ "N 1 #ˇ2
ˇ1 X 2 nk ˇˇ
ˇ N
PSD .k/ =dB D 10 log10 ˇ x .n/ w .n/ exp j ˇ 0k< :
ˇN N ˇ 2
nD0
(2.10)
After that, PSD estimate PSD .k/ is normalized to a SPL level of 96 dB, i.e., the
maximal is limited to 96 dB.
P .k/ =dB D 96 max fPSD .k/g C PSD .k/

(2.11)
D 4P C PSD .k/ ;
where 4P D 96 max fPSD .k/g. It is because we have no prior knowledge

regarding actual playback levels, the absolute pressure level of a sound can only
be specified by comparing to a reference. To this end, a sinusoid with amplitude
equal to half of PCM quantizer spacing A0 D 4 2
is defined as having a SPL of
0 dB, i.e., 20 log10 .A0 =A0 / D 0 dB. Consequently, for 16-bit
PCM data, a sinusoid
. 2 /
16 1 4
with amplitude equal to the overload level of quantizer Amax D 2
would
have a SPL of about 96 dB, i.e., 20 log10 .Amax =A0 / D 20 log10 216 1 96 dB
[51].
An example of the initial and normalized PSD estimates as well as the threshold
in quiet are shown in Fig. 2.19, where the frequencies of two graphs are plotted on
linear and Bark scales, respectively. Note that in psychoacoustic models, an offset
depending on the overall bit rate is employed for the threshold in quiet. It is equal to
12 dB for bit rates no less than 96 kbits=s and 0 dB for bit rates less than 96 kbits=s
per channel [77]. Sound tracks used in our experiments are of CD quality, whose
bit rates are normally greater than 96 kbits=s. Therefore, by comparing Fig. 2.19 to
Figs. 2.7 and 2.8, the threshold in quiet in Fig. 2.19 is shifted downward by 12 dB.
• STEP 2: Identification of tonal and nontonal maskers
On account of “asymmetry of masking,” it is required to discern frequency
components as tonal (i.e., sinusoidal) and nontonal (i.e., noise-like) maskers. Tonal
maskers are selected from local maxima of normalized PSD estimate, P .k/. A local
maxima refers to the maximum PSD within its two neighbors:
P .k/ P .k C 1/ and P .k/ P .k 1/ (2.12)

a
Initial PSD
100
Normalized PSD
Threshold in Quiet
50
−50
−100
−150
0 4000 8000 12000 16000 20000
Frequency (Hz)
b
Initial PSD
100
Normalized PSD
Threshold in Quiet
50
−50
−100
−150
1 3 5 7 9 11 13 15 17 19 21 23 25
Fig. 2.19 Initial and normalized PSD estimates (a) Frequency on linear scale (b) Frequency on
Bark scale
If the value of a local maxima is at least 7 dB greater than that of its neighboring
components within a certain Bark range Dk , such a maxima will be marked as a
tonal masker. All the tonal components comprise the “tonal” set, STM :
STM D fP .k/ j ŒP .k/ P .k ˙ Dk / 7 dBg ; (2.13)
where Dk varies with different frequency indices.11
8
< f˙2g ; 2 < k < 63 $ 2Fs
N
63F
N
s
kHz
Dk 2 f˙2; ˙3g ; 63 k < 127 $ 63Fs
N kHz
127Fs
:
: N
f˙2; ˙3; : : : ; ˙6g ; 127 k 250 $ 127Fs
N
250F
N
s
kHz
One point to note is that [77] did not specify the value of Dk for 251 k 256,
because the maskers within this range are already dominated by the threshold in
quiet (as seen in Fig. 2.19) and have no contribution to masking threshold. Actually,
it is the first criterion for decimation in Step 3.
As the effect of masking is additive in the logarithmic domain, the SPL of each
tonal component is calculated by
h P .k1/ P .k/ P .kC1/
i
PTM .k/ =dB D 10 log10 10 10 C 10 10 C 10 10 : (2.14)
In addition, the remaining components within each critical band12 are treated to be
nontonal. So we sum up their intensities as the SPL of a single nontonal masker for
each critical band, PNM :
X P .j /
PNM k =dB D 10 log10 10 10 8P .j / … STM ; (2.15)
j
where that k is the frequency index nearest to the geometric mean13 of each critical
band. Correspondingly, all the nontonal components are put into the “nontonal”
set, SNM .
11
The frequency edges are calculated based on the sampling frequency Fs .
12
Critical band boundaries vary with the Layer and sampling frequency. ISO/IEC IS 11172-3 [77]
has tabulated such parameters in Table D.2a–f. In our case, Table D.2b for Layer I at a sampling
frequency of 44.1 kHz is adopted.
!1=M
Y
M
13
The geometric mean of a data set Œa1 ; a2 ; : : : ; aM is defined as am . It is sometimes
mD1
!1=M " #
Y
M X
M
called the log-average, i.e., am D 10ˆ 1
M
log10 .am / .
mD1 mD1
a 120
Normalized power spectrum
Local maxima
100 Tonal components
Non−tonal components
Threshold in Quiet
80 Removed components
60
40
20
−20
0 4000 8000 12000 16000 20000
Frequency (Hz)
b 120
Local maxima
100 Tonal components
Non−tonal components TM
Threshold in Quiet TM 2
TM3
1
80
Removed components
60
16.55 17.08 17.57 18.01 Barks
40 0.49 Bark
20
−20
1 3 5 7 9 11 13 15 17 19 21 23 25
Fig. 2.20 Tonal and nontonal maskers (a) Frequency on a linear scale (b) Frequency on Bark scale
Tonal and nontonal maskers are denoted by pentagram and asterisk symbols in
Fig. 2.20, respectively. Particularly, the associated critical band for each masker is
indicated in the graph on Bark scale.
• STEP 3: Decimation of invalid tonal and nontonal maskers
On considering their possible contributions to masking threshold, the sets of tonal
and nontonal maskers are examined according to two criteria as follows:
One rule is that any tonal and nontonal maskers below the threshold in quiet
are removed. That is, only the maskers that satisfy Eq. (2.16) are retained, where
ATH .k/ is the SPL of threshold in quiet at frequency index k:
PTM; NM .k/ ATH .k/ : (2.16)
For example, one of each tonal and nontonal maskers between 24 and 25 Barks is
discarded, as shown in Fig. 2.20b.
The other rule is to simplify any group of maskers occurring within a distance
of 0.5 Bark: only the masker with the highest SPL is preserved and the rest are
eliminated.
PTM; NM .k/ D arg max PTM; NM .k C k0 / : (2.17)

k0 2Œ0:5;0:5
For example, two pairs of tonal maskers between 17 and 19 Barks, fTM1 ; TM2 g
and fTM2 ; TM3 g, are inspected. As shown in an enlarged drawing on the right of
Fig. 2.20b, the distance between fTM1 ; TM2 g is 0.49 Bark, and TM1 has a lower
SPL than TM2 . Therefore, TM2 is preserved, whereas TM1 is removed. Similarly,
we dispose of TM3 but retain TM2 for fTM2 ; TM3 g.
ˇ
ˇ Distance W17:57 17:08 D 0:49 Bark
TM2 fTM1 ; TM2 g ˇˇ
SPL W PTM1 < PTM2
ˇ
ˇ Distance W18:01 17:57 D 0:44 Bark
TM2 fTM2 ; TM3 g ˇˇ
SPL W PTM2 > PTM3
In Fig. 2.20, the invalid tonal and nontonal maskers being decimated are denoted by
a circle.
• STEP 4: Calculation of individual masking thresholds
After eliminating invalid maskers, individual masking threshold is computed for
each tonal and nontonal masker. An individual masking threshold L .j; i / refers to
the masker at frequency index j contributing to masking effect on the maskee at
frequency index i . It corresponds to L Œz .j / ; z .i /, where z .j / and z .i / are the
masker and maskee’s frequencies in Bark scale. In MPEG psychoacoustic models,
only a subset of samples over the whole spectrum are considered to be maskees and
involved in the calculation of global masking threshold. The number and frequencies
of maskees also depend on the Layer and sampling frequency, as tabulated from
Table D.1a–f in [77]. In our case, Table D.1b for Layer I at a sampling frequency of
44.1 kHz is adopted, where 106 maskees are taken into account.
The individual masking thresholds for tonal and nontonal maskers, LTM
Œz .j / ; z .i / and LNM Œz .j / ; z .i /, are calculated by
LTM Œz .j / ; z .i / =dB D PTM Œz .j / C 4TM Œz .j / C SF Œz .j / ; z .i / (2.18)

LNM Œz .j / ; z .i / =dB D PNM Œz .j / C 4NM Œz .j / C SF Œz .j / ; z .i / ; (2.19)
where PTM Œz .j / and PNM Œz .j / are the SPLs of tonal and nontonal maskers
at a Bark scale of z .j /, respectively. The term 4X is called masking index, an
offset between the excitation pattern and actual masking threshold. As mentioned
in Sect. 2.3.1, the excitation pattern needs to be shifted by an appropriate amount in
order to obtain the masking curve relative to the masker. Because tonal and nontonal
maskers have different masking capability, i.e., the noise is a better masker than pure
tone, the masking indices of tonal and nontonal maskers are defined separately as
follows [77]:
4TM Œz .j / D 6:025 0:275z .j / (2.20)

4NM Œz .j / D 2:025 0:175z .j / : (2.21)
The term SF Œz .j / ; z .i / is the spreading function discussed already in Sect. 2.4.1.1.

Psychoacoustic Model 1 employs spreading function in Eq. (2.8), rewritten in the
following expression:
8
ˆ
ˆ 17d z 0:4PX Œz .j / C 11; 3 d z < 1
ˆ
ˆ
<.0:4P Œz .j / C 6/ d z; 1 d z < 0
X
10 log10 SF .d z/ =dB D ;
ˆ
ˆ17d z;
ˆ 0 dz < 1
:̂
17d z C 0:15PX Œz .j / .d z 1/ ; 1 dz < 8
(2.22)
where d z is the distance from the maskee to masker, d z D z .i / z .j /,
as defined in Sect. 2.4.1.1. PX Œz .j / refers to PTM Œz .j / in the case of tonal
masker, otherwise PNM Œz .j / for nontonal masker. Notice that for reasons of
implementation complexity, the masking is no longer considered if d z < 3 Bark or
d z 8 Bark and thereby LTM Œz .j / ; z .i / and LNM Œz .j / ; z .i / are set to 1 dB
outside the above ranges [77].
Figure 2.21 shows the individual masking thresholds for both tonal and nontonal
maskers survived from the decimation.
• STEP 5: Calculation of global masking threshold
The global masking threshold is the combination of individual masking thresholds
and the threshold in quiet. Since the mixture of masking is additive, the global
masking threshold at frequency index i is calculated according to
2 3
ATH.i / X
NTM
LTM Œz.j /;z.i / X
N NM
LNM Œz.j /;z.i /
LG .i / =dB D 10 log10 410 10 C 10 10 C 10 10 5;
j D1 j D1
(2.23)
a
120
Tonal components
Masking thresholds for tonal components
100 Non−tonal components
Masking thresholds for non−tonal components
80
60
40
20
−20
0 4000 8000 12000 16000 20000
Frequency (Hz)
b 120
Tonal components
Masking thresholds for tonal components
100 Non−tonal components
Masking thresholds for non−tonal components
80
60
40
20
−20
1 3 5 7 9 11 13 15 17 19 21 23 25
Fig. 2.21 Individual masking thresholds (a) Frequency on linear scale (b) Frequency on Bark
scale
where ATH .i / is the SPL of threshold in quiet at frequency index i , NTM and NNM
are the number of tonal and nontonal maskers, and LTM Œz .j / ; W and LNM Œz .j / ; W
are their corresponding individual masking thresholds.
The global masking threshold is denoted by a bold dashed black line in

Fig. 2.22.
• STEP 6: Determination of the MMT
The MMT is derived from the global masking threshold. As mentioned in Step 4,
the global masking threshold LG is computed on only a subset of samples (here
106 samples) over the frequency spectrum, i.e., 1 i 106. Then these spectral
subsamples are mapped onto 32 uniform subbands, as shown in Fig. 2.23. Each
subband contains N=232
D 512=2
32
D 8 samples. Therefore, the minimum masking
level in the nth subband .1 n 32/ is determined by the following expression:
LMin .n/ =dB D min LG .i / ; (2.24)

fid .i/2subband n
where fid .i / is the frequency index corresponding to the i th subsample. After

spreading every LMin .n/ .1 n 32/ over its subband with 8 samples, we get
the MMT LMMT :
LMMT .m/ D LMin .n/ m D Œ8 .n 1/ C 1 W 8n: (2.25)
2.4.1.3 Comparison Between Psychoacoustic Model 1 and Model 2
The general idea of implementation on Psychoacoustic Model 2 is similar to Model

1. However, the concrete operations of calculating MMT in Psychoacoustic Model 2
are quite different from that of Model 1, as depicted in the following steps [78,81]:
• STEP 1: FFT analysis and calculation of complex spectrum
The input to Model 2 is a set of 1,024 samples, twice longer than 512-point frame
in Model 1. Before performing FFT, a Hanning window is applied as well.
• STEP 2: Definition of threshold calculation partitions and spreading function
The notion of “threshold calculation partitions” is a significant difference in Model
2. Instead of identifying the tonal and nontonal maskers in each critical band in
Model 1, Model 2 groups the frequency lines into so-called threshold calculation
partitions. Such partitions are also of nonlinear widths, but with finer frequency
resolution than critical band. Each partition has a width of either one FFT line
(at low frequencies) or 1/3 critical band (at high frequencies), whichever is wider
[78]. According to this criterion, there are 57 partitions at a sampling frequency of
44.1 kHz by calculation. The result complies with Table D.3b in [77].
The spreading function in Model 2 is described by Eq. (2.7) and one specific
spreading function is defined for each partition. Note that 10 log10 SF .d z/ in
Eq. (2.7) is level-independent and thereby suitable for alleviating the computational
burden of convolution in Step 4.
a
120
Tonal components
100 Normalized power spectrum
Threshold in Quiet
Global masking threshold
80 Minimum masking threshold
60
40
20
−20
0 4000 8000 12000 16000 20000
Frequency (Hz)
b
120
Tonal components
100 Normalized power spectrum
Threshold in Quiet
Global masking threshold
80 Minimum masking threshold
60
40
20
−20
1 3 5 7 9 11 13 15 17 19 21 23 25
Fig. 2.22 Global masking threshold and minimum masking threshold (a) Frequency on linear
scale (b) Frequency on Bark scale
Subband number, n
1 4 8 12 16 20 24 28 32
106
100
Subsamples over the spectrum, i
80
60
40
20
0
0 32 64 96 128 160 192 224 256
Frequency index k
Fig. 2.23 Mapping between spectral subsamples and subbands
• STEP 3: Calculation of unpredictability measure and weighted partition energy

Rather than selecting the relevant tonal and nontonal maskers in each critical band,
Model 2 introduces the property of unpredictability to describe how predictable
(tonal -like) the frequency component is. Unpredictability measure depends on the
magnitude and phase of complex spectrum. After weighting the energy of each
frequency line with unpredictability measure, we sum them up as the weighted
energy of each partition.
• STEP 4: Convolution of weighted partition energy and spreading function
As the behavior of simultaneous masking, the partition spreads its weighted
energy into the adjacent partitions. The overall masking effect is computed by the
convolution of spreading functions and weighted energy of each partition.
• STEP 5: Calculation of tonality index and SMR
Tonality index is a measure in Model 2, which is not used in Model 1. It denotes
the relative tonality of the maskers in each partition. The value of tonality index
is limited to the range of 0 (high unpredictability and noise-like) and 1 (low
unpredictability and tonal). Based on tonality index as well as an attenuation shift
factor between NMT and TMN, the SMR of each partition is calculated.
100
MMT from Model 1
MMT from Model 2
80
60
40
20
−10
0 4000 8000 12000 16000 20000
Frequency (Hz)
Fig. 2.24 Comparison of MMTs from Psychoacoustic Model 1 and 2
• STEP 6: Determination of the MMT

After obtaining SMR, the masking level of each partition is calculated by multi-
plying SMR to the inverse of signal energy, and then it spreads evenly over the
frequency line(s) within the partition. Finally, the MMT is determined by taking the
bigger value between the masking level and the threshold in quiet.
Figure 2.24 illustrates a comparison of the MMTs from Psychoacoustic Model
1 and 2. In view of the overall trend, MMT from Model 1 is analogous to that
from Model 2, although a bit less accurate at low frequencies. Generally, the
difference in masking effect of two psychoacoustic models is not evident [78]. On
the other hand, as the price of high precision, Model 2 involves more calculations
such as finer resolution of partitions, unpredictability measure, and the convolution
process. Consequently, it slows down the speed of execution, which is against the
requirement of audio watermarking. Therefore, we prefer Psychoacoustic Model 1
for our application.
2.4.2 Modelling the Effect of Nonsimultaneous Masking
In addition to simultaneous masking, the effect of nonsimultaneous masking is also

well exploited for developing perceptual models.
In [51], a time-sliding window is adopted in modelling the effect of nonsimulta-
neous masking. To resemble pre- and post-masking curves in Fig. 2.10, a weighting
0.5
Post−masking threshold
Input audio
0.25
Normalized amplitude
−0.25
−0.5
0 20 40 60 80 100 120
Time duration (ms)
Fig. 2.25 Modelling the effect of post-masking
function of time is designed to be in a shape of bulge: a larger weight on components

near the center of window, but gradual attenuation on components near the edges.
Generally, it is assumed that such temporal smoothing is applied to signal spectrum,
resulting in a smoothed output signal in time domain.
Different from [51], the modified envelope of input audio was used to approx-
imate the effect of post-masking in [71]. In particular, the estimated masking
curve increases with the envelope of signal and decays as an exponential function
e ˛t . The decay constant ˛ .˛ 0/ controls decaying rate as required, where
˛ D 1:2 103 in Fig. 2.25.
2.5 Summary
The ultimate aim of this chapter is to establish a psychoacoustic model that emulates
the HAS. Accordingly, audio watermarking techniques are able to analyze the host
audio signal in order to determine how the watermarks can be rendered as inaudible
as possible.
The chapter started with the physiology of the peripheral auditory system
including the outer, middle, and inner ears. The outer ear collects sound waves in
the air and channels them to interior parts of the ear; the middle ear transforms the
acoustical vibration of sound waves into mechanical vibration and passes them onto
the inner ear; the inner ear transduces mechanical energy into nerve impulses that
are transmitted to the brain. Then, some fundamental concepts of psychoacoustics
2.5 Summary 49
such as SPL, loudness, human hearing range, threshold in quiet, and critical
bandwidth were introduced. The notions of two types of auditory masking, i.e.,
simultaneous and nonsimultaneous masking, were also explained. In simultaneous
masking, it is noted that the masking ability of narrowband noise is superior to
pure tone. Based on the acquired knowledge, the ways of constructing the models
for simultaneous and nonsimultaneous masking effects are investigated respectively,
particularly simultaneous masking. After reviewing several models for the spreading
of masking, we described the details of implementing Psychoacoustic Model 1 in
ISO/MPEG standard, followed by a comparison with Model 2. On balance, two
psychoacoustic models have similar perceptual quality, but Model 2 requires more
computation than Model 1. Consequently, we adopted Psychoacoustic Model 1 in
the audio watermarking scheme we developed in this book.
Chapter 3
Audio Watermarking Techniques
In recent years, there has been considerable interest in the development of audio
watermarking techniques. To clarify the essential principles underlying a diversity
of sophisticated algorithms, this chapter gives an overview of basic methods for
audio watermarking, such as least significant bit (LSB) modification, phase coding,
spread spectrum watermarking, cepstrum domain watermarking, wavelet domain
watermarking, echo hiding, and histogram-based watermarking.
As the first step towards a full investigation into various approaches, we start
with the details of performance evaluation undertaken in this book, including the
parameters employed during perceptual quality assessment and robustness tests.
Then, different audio watermarking techniques are separately implemented and
evaluated in order to ascertain their advantages and disadvantages. Also, possible
enhancements are exploited to further improve their capabilities. Finally, the chapter
is concluded with a summary of comparative study.
3.1 Specifications on Performance Evaluation
As discussed in Sect. 1.3.2, performance evaluation of audio watermarking systems

involves three major aspects, i.e., perceptual quality assessment, robustness test,
and security analysis. To ensure a fair comparison of the different techniques, this
section states the details of evaluation methods and test parameters. Without further
notice, these specifications apply to all the performance evaluations hereafter,
including the experiments carried out in this chapter and Chap. 5.
52 3 Audio Watermarking Techniques
Left channel of stereo signal

0.7
0
−0.7
0 4 8 12 16 20 24
Right channel of stereo signal
0.7
−0.7
0 4 8 12 16 20 24
Time (s)
Fig. 3.1 An example of a two-channel stereo signal
3.1.1 Audio Test Signals Used for Evaluation
A collection of audio test files are prepared for performance evaluations in this
book. The test set contains seventeen pieces of audio signals in total and all of
them are in WAVE format (44.1 kHz, 16 bit, mono). For simplicity of reference,
each audio signal An is marked with a subscript number n, i.e., (i) Vocal: Soprano1 ,
Bass2 , Quartet3 ; (ii) Percussive instruments: Hihat4 , Castanets5 , Glockenspiel16 ,
Glockenspiel27 ; (iii) Tonal instruments: Harpsichord8 , Violoncello9 , Horn10 ,
Pipes11 , Trumpt12 , Electronic tune13 ; (iv) Music: Bach14 , Pop15 , Rock16 , Jazz17 .
More details of each audio test file, such as its duration and waveform, are listed in
Appendix E.
Except for the category of music, A14 A17 , the majority of test files are
selected from audio tracks on the EBU SQAM1 disc, specifically for the testing
and evaluation of sound systems [79]. Originally, most audio tracks are stereo
channels. Since our study is not limited to stereo audio watermarking,2 , the left
channel of each signal is always used for the watermarking. Figure 3.1 shows an
example of a stereo signal and the left channel signal is taken as the audio test
file A2 . It is observed that there are silent intervals inherent in audio data. Long
silence is usually intractable in the watermarking, because embedding watermarks
in muteness would unavoidably introduce perceived noise. Thus, watermarking
1
EBU: The European Broadcasting Union; SQAM: Sound Quality Assessment Material.
2
In general, stereo audio watermarking depends on some kind of relation between two channels
[80], so it can only apply to stereo signals. However, mono audio watermarking can commonly
treat one stereo channel as two mono channels, so it supports both mono and stereo audio signals.
3.1 Specifications on Performance Evaluation 53
regions must be carefully chosen. This issue will be further elaborated in Chap. 4.
To avoid long silence during the watermarking, we only embed the watermark into
the first half of A2 for performance evaluations in this chapter.
3.1.2 Implementation of Perceptual Quality Assessment
As mentioned in Sect. 1.3.2.1, perceptual quality assessment on audio watermarking

is comprised of subjective listening tests and objective evaluation tests.
Subjective listening tests are carried out in an isolated chamber, where ten trained
listeners with different audio engineering backgrounds are participants. All the
stimuli are presented through a high-fidelity headphone. In the MUSHRA test, the
participants are asked to rate the perceptual quality of each watermarked signal
relative to its host signal using a MATLAB’s graphical user interface (GUI). The
details of the developed GUI will be described in Sect. 5.2.1. Also, a rating based on
the five-scale subjective difference grade (SDG) is performed as well. According to
the descriptions in Table 1.2, the watermarked audio signals are expected to possess
SDGs between 1:0 and 0. It is also acceptable if SDGs are less than 2:0 [2].
Moreover, software PEAQ [48] is utilized to provide an objective difference grade
(ODG). ODG is an objective measurement of SDG and its specifications conform to
those of SDG that are described in Table 1.2. The reason for choosing PEAQ rather
than EAQUAL and PEMO-Q is that PEAQ is an improved version of EAQUAL and
free to use, while PEMO-Q is a commercial software tool and its demo version is
restricted to signal lengths up to 4 sec only. In addition, the signal-to-noise ratio
(SNR) defined in Eq. (1.3) is calculated as an objective indicator of perceptual
quality.
3.1.3 Implementation of Robustness Test
The general guideline for robustness tests was discussed in Sect. 1.3.2.2. On
consideration of test items in SDMI, STEP2000, and StirMark for Audio, two
robustness tests are set up in the book, i.e., a basic robustness test and an advanced
robustness test.
Recall that in robustness tests, the attacked watermarked signals should not
be degraded far beyond tolerable levels. On the basis of this premise, the attack
parameters listed below are determined accordingly.
3.1.3.1 Basic Robustness Test
The basic robustness test incorporates a variety of typical attacks on audio water-
marking techniques. Table F.1 in Appendix F shows the parameters used, expres-
sion, and implementation of each attack. The basic robustness test is employed for
evaluating different audio watermarking techniques in Sect. 3.2 below. It is worth

mentioning that the software “Adobe Audition v3.0”3 is used to implement some
attacks.
Noise addition: Add white Gaussian noise to the watermarked audio signal to reach
the specified SNR. Usually, the SNR is targeted at a value between 20 and 40 dB.
Resampling: The watermarked audio signal that originally has a sampling rate
of 44.1 kHz is downsampled to 22.05 or 11.025 kHz and then upsampled back to
44.1 kHz.
Requantization: The watermarked audio signal that originally has 16 bits/sample is
requantized down to 8 bits/sample and then requantized back to 16 bits/sample.
Amplitude scaling: The amplitude of the watermarked audio signal is rescaled by
˙10 % or ˙20 %. A positive and negative rate of scaling denotes that the amplitude
is amplified and attenuated, respectively.
Low-pass filtering: A low-pass filter with a cutoff frequency of 4, 6, or 8 kHz is
applied to the watermarked audio signal.
Echo addition: An echo with a delay of 100 or 200 ms and a decay of 20 % or 30 %
is added to the watermarked audio signal.
Reverberation: Reverberation in a large empty hall with a reverberation time4 of
1 s is exerted on the watermarked audio signal.
MP3 compression: The watermarked audio signal originally in .wav format is
compressed at a bitrate of 48, 64, 96, or 128 kbps (kilobits per second) by an
MP3 encoder. Then the .mp3 file is decompressed back to .wav format by the MP3
decoder. One point to note is that the process of compression/decompression causes
not only amplitude modification but also displacement between the watermarked
and attacked audio signals. A certain amount of quasi-zero samples are padded at the
inception5 and the end of attacked signal [81], because of internal data organization
in MP3 files. Therefore, MP3 compression has two forms, i.e., Compression I
and Compression II. In Compression I, we cut off those extra samples so as to
focus on the effect of amplitude modification by MP3 compression. Meanwhile,
Compression II is actually a combined attack that combines data compression and
zeros inserting.
3
Adobe Audition is a powerful digital audio recorder, editor, and mixer for Windows. It can
perform a lot of operations, such as resampling, requantization, amplitude scaling, reverberation,
MPEG compression, time stretching, and pitch shifting, on various formats of audio files, .au, .voc,
.vox, .wav, and so on.
4
The reverberation time of a room is the time that it takes for sound to decay by a certain level ˛
dB once the source of sound has stopped [30]. T60 is when ˛ D 60 dB.
5
Based on extensive operations with Adobe Audition v3.0, it is found that an amount of 1,201
samples is added to the beginning of an audio file.
3.1 Specifications on Performance Evaluation 55
DA/AD conversion: The watermarked audio signal is played through the audio
player in a computer. Then the playback signal is recorded by connecting the
headphone jack to the line-in jack on the sound card of the computer.
Random samples cropping: A number of 25 ms intervals are cropped at randomly
selected positions in the front, middle, and rear of the watermarked audio signal.
Jittering: Jittering is an evenly performed form of random samples cropping. For
our watermarked audio signal, 0.1–0.2 ms out of every 20 ms is cropped.
Zeros inserting: A number of 25 ms silent intervals are inserted into randomly
selected positions in the front, middle, and rear of the watermarked audio signal.
Pitch-invariant time-scale modification (PITSM): The time-scale of the water-
marked audio signal is stretched from ˙4 % up to ˙10 %, whereas the audio pitch
is preserved. Positive PITSM results in a longer duration with a slower tempo, while
negative PITSM results in a shorter duration with a faster tempo.
Tempo-preserved pitch-scale modification (TPPSM): The pitch-scale of the water-
marked audio signal is shifted from ˙4 % up to ˙10 %, whereas the audio tempo is
preserved. Positive TPPSM results in a higher pitch, while negative TPPSM results
in a lower pitch.
The last five attacks belong to desynchronization attacks, which cause displace-
ment between the encoder and decoder. Therefore, it is difficult to retrieve a
watermark suffering from such hazardous attacks, especially PITSM and TPPSM.
3.1.3.2 Advanced Robustness Test
The advanced robustness test involves more stringent attacks than the basic robust-
ness test and is specifically designed for rigorously evaluating our proposed audio
watermarking algorithm to be described in Chap. 4. It consists of three parts: a
test with StirMark for Audio, a test under collusion, and a test under multiple
watermarking.
Collusion: We separately embed n different watermarks w.1/ .2/ .n/
o ; wo ; : : : ; wo into a
.1/ .2/ .n/
host signal so and obtain n watermarked signals sw ; sw ; : : : ; sw correspondingly.
Without loss of generality, these watermarked signals are further combined to create
n average watermarked signals sw.i/ .1 i n/ as follows:
8
<sw.j / D Embedding s ; w.j / ; 1 j n;
o o
:s .i/ D 1 s .1/ C s .2/ C C s .i/ ; 1 i n:
(3.1)
w i w w w
In the detection, i watermarks w.i;/ e are detected from the average watermarked
signal sw.i/ individually:

w.i;j
e
/
D Detection sw.i/ ; 1 i n and 1 j i; (3.2)
where n D 2 4 in our robustness test.

Multiple watermarking: n different watermarks wo.1/ ; wo.2/ ; : : : ; w.n/

o are sequen-
tially embedded in the following way:
(
sw.1/ D Embedding so ; w.1/
o ;
(3.3)
sw D Embedding sw ; wo ; 2 i n:
.i/ .i1/ .i/
In the detection, i watermarks are detected from the watermarked signal sw.i/
individually:

w.i;j
e
/
D Detection sw.i/ ; 1 i n and 1 j i; (3.4)
where n D 2 4 in our robustness test.

In multiple watermarking, one point to note is that the technique for each
watermarking may be the same or different. For example, the first watermark
is embedded using our proposed algorithm, but the second watermark may be
embedded using echo hiding watermarking. In this case, the corresponding detection
method is employed to detect each watermark.
3.2 Audio Watermarking Algorithms
Over the last years, many digital watermarking methods have been proposed for dif-
ferent applications. These methods can be broadly divided into two main categories:
(1) blind embedding, where the encoder does not exploit the knowledge of the host
signal, for example, spread spectrum watermarking, and (2) informed embedding,
where the knowledge of the host signal is adequately exploited by the encoder,
for example, quantization index modulation (QIM) [1, 82].6 Both prototypes have
found implementations in audio watermarking, such as LSB modification, phase
coding, spread spectrum watermarking, cepstrum domain watermarking, wavelet
domain watermarking, echo hiding, and histogram-based watermarking. In the fol-
lowing subsections, one algorithm of each technique is implemented and evaluated
separately.
Regarding the performance evaluations in this section, perceptual quality assess-
ment and robustness tests are conducted as follows. In the perceptual quality
assessment, the ODG using software PEAQ and the SNR are employed to indicate
the audio quality. Since subjective listening tests are costly and time-consuming,
only informal subjective listening tests are carried out if necessary. Informal
subjective listening tests are performed in the same environment as described in
Sect. 3.1.2; however, only a couple of listeners are involved. Without providing
SDG scores, they are merely required to ascertain whether the watermarked signal
is perceptually undistinguished from the host signal.
6
Two categories are named as the host-interference nonrejecting method and the host-interference
rejecting method respectively in [35, 83, 84].
3.2 Audio Watermarking Algorithms 57
For the robustness test, some or all the attacks in basic robustness test are
involved and the bit error rates (BERs) as defined in Eq. (1.4) are calculated
accordingly. Moreover, repetition coding on the watermark is adopted to enhance
the robustness. For a .nr ; 1/ repetition code, each watermark bit is repeated nr
times and subsequently embedded. In the detection, the bits are determined using
the majority vote rule. For example, repetition coding
with nr D 3 on the original
watermark wo D 1 0 1 yields a sequence w O o D 111 000 111 . Suppose that the

attacked sequence becomes wO e D 011 001 110 , then the final detected watermark

is we D 1 0 1 and BER = 0 %.
3.2.1 Least Significant Bit Modification
Earlier audio watermarking techniques embed the watermark into the host signal in
a straightforward manner. One method is to replace the LSB of each sample with the
watermark represented in a coded binary string [36]. In this way, the data payload
of the watermarking system could be very high, approximately of the same order of
magnitude with the sampling frequency of the host signal. Ideally, for instance, the
bit rate is 44.1 kbps for an audio signal with a sampling frequency of 44.1 kHz [31].
However, the system under such conditions would be quite irresistible to any attack.
In order to enhance the robustness and security, LSB modification can be performed
on some selected subsets of the samples only, such as low-frequency components
that are perceptually important. Usually, repetition coding could help increase the
detection rate in LSB watermarking.
3.2.1.1 Algorithm
The audio watermarking technique in [85] is an example of the improved LSB

modification, developed on the basis of the method in [86]. A host signal is first
decomposed by a L-level complementary filterbank. Then the output from the low-
pass filter is scrambled by a pseudorandom sequence (PRS) to increase the security.
After applying the modified discrete cosine transform (MDCT), the coefficients at
different orders are quantized to embed the watermark. In MATLAB, sound waves
are normalized to a magnitude of one. Therefore, the coefficients are rounded at
different decimal levels in the implementation of [85]. Moreover, repetition coding
with nr D 4 is used to reduce bit errors in the detection. Typically, it is a kind of
LSB modification when the last decimal number is altered.
a
0.7
0
−0.7
0 2 4 6 8 10
b
0.7
0
−0.7
0 2 4 6 8 10
c
0.01
0
−0.01
0 2 4 6 8 10
Time (s)
Fig. 3.2 Host signal and a watermarked signal by LSB modification. Note that the watermarked
signal is produced by using L D 6 and modifying the third and fourth decimal places. (a) Host
audio signal. (b) Watermarked audio signal. (c) Difference between the watermarked and host
audio signals
3.2.1.2 Performance Evaluation
The performance of the improved LSB modification is evaluated in this section.

Figure 3.2 shows the host signal and a watermarked signal by the improved
LSB modification. Also, the results of performance evaluation are summarized in
Table 3.1. As illustrated in Fig. 3.2c, the difference between the watermarked and
host audio signals is small, which corresponds to a high SNR (i.e., 52.17 dB) in
Table 3.1.7
From Table 3.1, the determination of decimal places results in a compromise
between imperceptibility and robustness. On one hand, it is better to modify
insignificant bits for the good of imperceptibility, such as the third or fourth decimal
places. On the other hand, it is more robust if significant bits are used for embedding
the watermark, such as the first or second decimal places. Moreover, the selection of
a complementary filterbank level controls the balance between imperceptibility and
data payload. By employing the complementary filterbank with less levels, we can
embed more watermarking bits; nevertheless, such acquirements are at the expense
of the degradation of the SNR.
7
Note that two of the ODGs in Table 3.1 are slightly positive, i.e., 0.13 and 0.03. According to its
definition, the ODG should normally be in the range [ 4, 0]. However, if the distortion caused
by watermarking is very low, then the cognitive model calculates positive values. In such cases, it
is interpreted that the distortion is mostly inaudible for humans [38].
Table 3.1 Results of performance evaluation of LSB modification

Watermarking parameters
Decimal place 1st and 2nd 2nd and 3rd 3rd and 4th
Filterbank level 8 10 4 6 8 10 4 6
Watermark length Nw 512 128 8192 2048 512 128 8192 2048
(1) Perceptual quality assessment
SNR/dB 17:19 23:75 25:98 31:96 38:00 44:28 45:96 52:17
ODG 3:18 1:25 3:01 1:13 0:32 0:13 0:96 0:03
(2) Robustness test (BER: %)
No attack 0 0 0 0 0 0 0 0
Noise (36 dB) 0 0 0:98 1:03 0:59 3:13 22:13 23:29
Lp filtering (8 kHz) 0:78 0:78 24:26 27:78 17:38 20:31 46:36 47:46
Echo (0.3, 100 ms) 13:28 15:63 42:44 47:27 41:41 42:97 48:56 49:71
Compression I (96 kbps) 0:39 0:78 10:88 13:43 11:33 14:06 43:69 45:17
Compression II (96 kbps) 46:68 39:06 49:29 48:83 49:61 40:63 49:74 50:73
On the whole, the watermarked signals produced by rounding off different

decimal places are robust against the attacks to some extent, except that the
detections fail completely under a desynchronization attack—Compression II.
However, LSB modification in the strict sense would survive only in closed, digital-
to-digital environments.
3.2.2 Phase Coding
Based on the fact that the human auditory system (HAS) is unable to perceive
the absolute phase, only the relative phase [36], audio watermarking techniques
can embed watermarks into the phase of host signal, i.e., phase coding and phase
modulation [31].
3.2.2.1 Algorithm
The basic phase-coding method was presented in [36]. It splits the host signal into
frames and the first frame’s phase spectrum is modified to represent the watermark.
Then the phases of subsequent frames are changed accordingly to preserve their
relative phases. Thus, the first frame is crucial for watermark embedding and it
must not be an absolute silence. For watermark detection, the first frame of the
attacked signal is taken out, and then the value of its phase spectrum is calculated
to determine the watermark. The premise for detection is precise synchronization to
obtain the first frame accurately.
Algorithm 3.1 describes the pseudocode of phase coding. Note that the length of
each˘ frame is N , but the watermark has a length of .N 2 1/ only, where N 2 D
N
2
and bc is the smallest integer value. The reason is that the Fourier transform of
a real-valued signal exhibits conjugate symmetry [87] and only half of the spectrum
is available to embed the watermark. Moreover, the first spectral component is DC
value whose phase is always equal to 0 or , so it is not used for watermarking.
MATLAB script for phase-coding method can be found as Phase coding:m file
under Audio_Watermarking_Techniques folder in the attached CD.
Table 3.2 shows the results of performance evaluation of the basic phase-coding
method. As indicated by very low SNRs, the watermarked signals have been
changed greatly and their perceptual quality has degraded badly.
The cause of deterioration is the substitution of ˆ1 with a binary sequence of 2
and 2 . In the basic method, every component of the phase spectrum is altered to
represent the watermark. Such a sharp phase transition is likely to produce audible
distortion [36]. In order to smooth the variation, the modified method is to change
the phase spectrum at an interval of ne and perform interpolation between the
values. Several kinds of interpolation such as linear interpolation and cubic spline
interpolation were tested.
Figure 3.3 illustrates an example of the watermarked signal produced by the
modified phase-coding method. From Fig. 3.3c, there is still quite a difference
between the watermarked and host signals, which also interprets very low SNRs
in Table 3.2. Informal subjective listening tests show that the perceptual quality
of the watermarked signals is not satisfied yet. Moreover, as shown in Table 3.2,
the watermarked signals by phase coding are not robust and nearly all the BERs
are over 20 %. In addition, repetition coding is not helpful to phase coding. This is
because the effect of nr times repetition coding resembles that of embedding interval
ne D nr with linear interpolation.
To achieve imperceptible watermarking, Kuo et al. [88] proposed phase modula-
tion under the following constraint condition:
ˇ ˇ
ˇ d .z/ ˇ
ˇ ˇ ı
ˇ d z ˇ < 30 ; (3.5)
where .z/ denotes the signal phase and z is the Bark scale [51]. More information
on phase modulation for audio watermarking can be found in [31, 80, 81].
Algorithm 3.1 Pseudocode of basic phase coding method

% Watermark embedding
% Note: Host signal so .n/, 1 ˘ n No is split into NP frames fgi g with N samples,
% where NP D NNo .
sw D Œ ; % Initialize the watermarked signal sw .n/
so .n/ D fgi .j /g, 1 i NP and 1 j N ; ˘
Binary watermark wo .m/, 1 m .N 2 1/; % N 2 D N
2
% Loop through every frame gi

for i D 1 W NP
% Calculate the amplitude spectrum i and phase spectrum î
Œi ; î D F .gi /; % F ./ is the Fast Fourier Transform (FFT).
if i D 1
‰i .1/ D î .1/;
% Embed the watermark into the phase of the first frame
for m D 1 W .N 2 1/
if wo .m/ D 1, then ‰i .m C 1/ D 2 ;
if wo .m/ D 0, then ‰i .m C 1/ D 2 ;
end
else
4î 1 D î .1 W N 2/ î 1 .1 W N 2/;
‰i D ‰i 1 C 4î 1 ;
end
f
i D exp .j‰i /; % The first half of the phase spectrum
% Function f liplr ./ is to flip a matrix left to right.
% Function conj ./ is to compute
the complex conjugate.
f
s
i D f liplr conj i ; % The second half of the spectrum
h i
f
i D i .1 W N 2 1/ ;
s
i ; 0;
% Reconstruct the watermarked frame giw
giw D F 1 .i ; i /; % F 1 ./ is the Inverse Fast Fourier Transform (IFFT).
sw D sw ; giw .
end
% Watermark detection
% Note: sa is the attacked signal.
1 aD saa .1 W N /; %
a
g The first frame
1 ; ˆ1 D F g1a ;
% The watermark we is detected based on the phase spectrum â1 .
for m D 1 W .N 2 1/
if â1 .m/ 0, then we .m/ D 1;
if â1 .m/ < 0, then we .m/ D 0.
end
Table 3.2 Results of performance evaluation of phase coding
Phase-coding method Basic method Modified method
Frame length N 1;024 2;048 4;096 2;048 2;048 4;096 4;096
Embedding interval ne 1 1 1 128 64 128 64
Watermark length Nw 511 1; 023 2; 047 8 16 16 32
SNR /dB 2:55 2:53 2:62 0:23 0:26 2:15 1:35

No attack 0 0 0 0 0 0 0
Noise (36 dB) 19:57 29:03 40:35 12:50 18:75 31:25 43:75
Lp filtering (8 kHz) 34:25 37:44 35:66 25:00 25:00 37:50 21:88
Compression I (96 kbps) 18:98 21:02 23:69 37:50 50:00 37:50 59:38
Compression II (96 kbps) 54:21 44:97 49:73 87:50 50:00 50:00 37:50
a
0.7
−0.7
0 2 4 6 8 10
b
0.7
−0.7
0 2 4 6 8 10
c
0.7
−0.7
0 2 4 6 8 10
Time (s)
Fig. 3.3 Host signal and a watermarked signal by the modified phase-coding method. Note that
the watermarked signal is produced by watermarking with N D 2; 048 and ne D 128. (a) Host
audio signal. (b) Watermarked audio signal. (c) Difference between the watermarked and host
audio signals
a
Host signal Watermarked signal
s sw
Psychoacoustic Scaling factor
model
Watermark Watermark wm
wo spreading Modulated watermark
Pseudorandom sequence
rs
b
Attcked signal ~
sa Extracted watermark
Pre-processing Correlator we
sa
Pseudorandom sequence
rs
Fig. 3.4 Block diagram of basic SS watermarking scheme. (a) Embedding process. (b) Detection
process
3.2.3 Spread Spectrum Watermarking
Spread spectrum (SS) watermarking is considered to be the most popular technique

for digital watermarking [4, 31]. It spreads the watermark throughout the spectrum
of the host signal, so that the signal energy present in every frequency bin is very
small and hardly detectable. In this way, the embedded watermark can possess
a large measure of security as well as robustness [89]. However, the process
of watermarking may easily introduce perceivable distortion to audio files. So
amplitude shaping by the masking threshold from the psychoacoustic model is often
employed to keep the watermark inaudible [1, 4, 90].
3.2.3.1 Algorithm
There are two main forms of SS watermarking, namely direct sequence spread
spectrum (DSSS) [8, 89, 91] and frequency hopping spread spectrum (FHSS) [92].
DSSS-based audio watermarking method is more commonly used and its basic
scheme is shown in Fig. 3.4 [1,90]. In watermark embedding process, the watermark
wo is modulated by PRS rs to produce the modulated watermark wm . To keep wm
inaudible, scaling factor ˛ may be used to control the amplitude of wm . Then the
watermarked signal sw is produced by adding wm to the host signal so . In watermark
detection process, the watermark we is extracted by correlating the received signal
sa with the PRS rs used in the embedding.
Note that the watermark can be spread not only in the time domain but also in
various transformed domains. Discrete Fourier transform (DFT), discrete cosine
transform (DCT), and discrete wavelet transform (DWT) are some examples of
transforms that are frequently used. Typically, a SS watermarking scheme that
spreads the watermark into the time-domain signal is implemented as follows.
First, host signal˘ so .n/ ; 1 n No is split into NP frames fgi g with N samples,
where NP D NNo :
so .n/ D fgi .j /g ; 1 i NP and 1 j N: (3.6)
So a number of NP watermark bits can be embedded at most, i.e., Nw D NP .
Each watermark bit wo .i / 2 fC1; 1g is modulated by one PRS rs . A sequence of
random numbers uniformly distributed in the interval .0:5; 0:5/ is applied in our
experiment. Then, the watermarked frame giw is obtained by adding the modulated
frame to the host frame as follows:
giw D gi C ˛wo .i / rs ; (3.7)
where the factor ˛ controls the strength of watermarking.
For a better perceptual quality, adaptive factor ˛ D ˇ max .abs
˚ .gi // is adopted,
where ˇ is a scaling factor. Finally, all the watermarked frames giw ; 1 i Nw
are concatenated to produce the watermarked signal sw .
In the detection, watermark bits are determined by using a linear correlation
between the watermarked
˚ signal and PRS. After splitting the watermarked signal
sw .n/ into frames giw in the same way as the embedding, the linear correlation
Rc ./ between each frame giw and rs is calculated as
1 X w
N
Rc .i / D g .j / rs .j /
N j D1 i
1 X
N
D Œgi .j / C ˛wo .i / rs .j / rs .j /
N j D1
1 X 1 X
N N
D gi .j / rs .j / C ˛wo .i / Œrs .j /2 : (3.8)
N j D1 N j D1
Ideally, if the host frame gi and the PRS rs are independent, the first term in Eq. (3.8)
is close to zero. Meanwhile, the second term has a large magnitude and its sign
depends on the watermark bit wo .i /. However, it is not always the condition that
gi and rs are uncorrelated. In this case, the first term has similar or even larger
magnitude than the second term, which would lead to incorrect detection. Thus, the
watermarked signal must be preprocessed to reduce the effect of the host signal to
the fullest extent [90]. To this end, different preprocessing methods are developed,
such as linear predictive coding (LPC) filtering [93], cepstrum filtering [8], the
Savitzky–Golay filtering [4], and decorrelation by subtracting the host signal [89]
or adjacent frames [9].
For simplicity of implementation here, we assume the host signal is known at

the decoder, i.e., a non-blind watermarking. Then, gi is subtracted directly from
giw to eliminate the first item in Eq. (3.8). After that, each watermark bit we .i / is
determined based on the correlation value Rc .i /:
8
<1; if Rc .i / 0;
we .i / D (3.9)
:0; if R .i / < 0:
c
MATLAB script for SS watermarking can be found as SS watermarking:m file

under Audio_Watermarking_Techniques folder in the attached CD.
Figure 3.5 shows the host signal and a watermarked signal by SS watermarking. As a
result of a small difference in Fig. 3.5c, the SNRs in Table 3.3 are higher than 30 dB.
For perceptual quality assessment, however, the values of the ODG are pretty low.
To ascertain the real perceptual quality, we carry out an informal subjective listening
test on the watermarked signal with the highest ODG in Table 3.3, i.e., the shaded
column. When the watermarked signal is played at a high volume, a constant hissing
background noise is heard. This occurs as a result of the modulated PRS added to
the host signal as white noise. Therefore, amplitude shaping by the psychoacoustic
model is very important to produce an unperceived watermark.
In the robustness test, SS watermarking with repetition coding behaves differ-
ently. In descending order of resistance, the attacks are sorted as Compression I,
noise addition, requantization, low-pass filtering, echo addition, resampling, and
Compression II. Apparently, SS watermarking is quite vulnerable to desynchroniza-
tion attacks. Thus, proper solutions for the synchronization problem are necessary
to improve the detection rate.
3.2.4 Cepstrum Domain Watermarking
Cepstrum domain watermarking is performed to embed the watermark into cepstral

coefficients. By definition, the complex cepstrum xO .n/ is the inverse Fourier
transform of the complex logarithm of the Fourier transform of a signal x .n/.
Mathematically, it is described as follows [94]:
X
N 1
X e j!
D x .n/ e j!n
nD0
˚ ˇ ˇ
XO e j! D log X e j! D log ˇX e j! ˇ C j arg X e j! (3.10)
Z
1
xO .n/ D XO e j! e j!n d!;
2

a
0.7
−0.7
0 2 4 6 8 10
b
0.7
−0.7
0 2 4 6 8 10
c
0.01
−0.01
0 2 4 6 8 10
Time (s)
Fig. 3.5 Host signal and a watermarked signal by SS watermarking. Note that the watermarked
signal is produced by watermarking with N D 4; 096, nr D 3 and ˇ D 0:03. (a) Host audio signal.
(b) Watermarked audio signal. (c) Difference between the watermarked and host audio signals
where log ./ refers to the natural logarithm. In a more compact way, xO .n/ D
F1 flog . F fx .n/g/g, where F ./ is the Fourier transform and F1 ./ is its inverse.
The complex cepstrum xO .n/ has preserved information about the magnitude and
phase of the frequency spectrum of x .n/. Therefore, x .n/ can be recovered from
xO .n/ by the inverse complex cepstrum. Figure 3.6 shows the block diagram of
computing the complex cepstrum and the inverse complex cepstrum using DFT and
IDFT [87].
In MATLAB implementation, the real part of the complex cepstrum is often
utilized as the common “cepstrum” c .n/ [95], i.e.,
c .n/ D real . xO .n/ / : (3.11)
Note that the real part of the complex cepstrum c .n/ should be distinguished from
the “real cepstrum,” cr .n/. The real cepstrum is defined as the inverse Fourier
transform of the logarithm of the magnitude of the Fourier transform of a signal
x .n/, i.e.,
Z ˇ ˇ
1 ˇ O j! ˇ j!n
cr .n/ D ˇX e ˇe d!: (3.12)
2

Table 3.3 Results of performance evaluation of SS watermarking
Frame length N 2048 4096
Repetition coding nr 1 3 1 3 1 3 1 3
Watermark length Nw 256 85 256 85 128 42 128 42
Scaling factor ˇ 0:02 0:02 0:03 0:03 0:02 0:02 0:03 0:03
SNR /dB 34:81 34:59 31:09 31:11 33:90 33:92 30:46 30:46
ODG 3:68 3:67 3:76 3:73 3:70 3:68 3:77 3:78
No attack 0 0 0 0 0 0 0 0
Noise (30 dB) 7:03 4:71 3:52 1:18 3:13 2:38 0:78 0
Re-sampling (22.05 kHz) 46:88 34:12 31:25 28:24 30:47 28:57 17:97 14:29
Requantization (8 bit) 4:69 5:88 2:34 0 4:69 0 1:56 0
Lp filtering (8 kHz) 12:89 11:76 4:69 0 7:03 0 0:78 0
Echo (0.3, 100 ms) 28:52 10:59 10:94 2:35 7:03 7:14 2:34 0
Compression I (96 kbps) 1:17 0 0:39 0 0 0 0 0
Compression II (96 kbps) 50:00 47:06 49:61 58:82 52:34 45:24 48:44 47:62
a
Time domain Cepstrum domain
Discrete Complex Inverse discrete
x(n) Fourier transform logarithm Fourier transform xˆ (n)
DFT{.} log{.} IDFT{.}
b
Cepstrum domain Time domain
Discrete Complex Inverse discrete
xˆ (n) Fourier transform natural exponential Fourier transform x(n)
DFT{.} exp{.} IDFT{.}
Fig. 3.6 Block diagram of computing the complex cepstrum and the inverse complex cepstrum.
(a) Complex cepstrum xO .n/ D F1 flog . F fx .n/g/g. (b) Inverse complex cepstrum x .n/ D
F1 fexp . F fxO .n/g/g
Different from complex cepstrum, real cepstrum has lost the phase information.
Therefore, x .n/ cannot be reconstructed perfectly from cr .n/.
3.2.4.1 Algorithm
There are some audio watermarking schemes conducting in the cepstrum domain
[95–99]. Li et al. [95] found that the statistical mean of cepstrum coefficients is an
attack-invariant feature and accordingly developed cepstrum domain watermarking
based on statistical mean manipulation (SMM). In this way, the statistical mean of
cepstrum coefficients, , is set to be ˙˛w that represent watermark bit “1” and “0”
respectively. Apparently, the strength of watermarking is controlled by ˛w . For the
detection, the statistical mean of watermarked cepstrum coefficients w is calculated
and compared to a predefined threshold Td . If w Td , then the watermark bit
is “1.” Otherwise, it is “0.” Note that instead of the mean, we use the sum for
bit determination in practice. This is because the mean of cepstrum coefficients is
usually around 104 , which is rather small for a comparison.
Algorithm 3.2 describes the pseudocode of the watermarking process.
Furthermore, Hsieh et al. [97] proposed a method of embedding based on time
energy features to solve the synchronization problem. The watermark is embedded
into the frames followed by salient points, the positions where signal energy
climbs fast to a peak [33]. As salient points are supposed to remain stable after
attacks, synchronization can be regained in the detection. Afterward, Cui et al. [98]
improved the method in [95] by employing the psychoacoustic model to control the
audibility of the introduced distortion. Apart from SMM, Gopalan [99] embedded
the watermark by altering the cepstrum in the regions that are psychoacoustically
masked, so as to ensure a better trade-off between imperceptibility and robustness.
Different from [95, 97–99], Lee et al. [96] spread the PRS in the cepstrum
domain to watermark the audio signal, considered in some sense as kind of a
SS watermarking. Also, in order to minimize its audibility, the PRS is weighted
according to the distribution of the cepstrum coefficients and the masking threshold
from the psychoacoustic model to minimize its audibility. But the scheme is non-
blind because the host signal is required in the detection.
3.2.4.2 Strategies for Improvement
Based on the algorithm in [95], we implement improved cepstrum domain water-

marking. Several strategies are used to enhance the system performance. First of all,
[95] did not mention the smooth transition between adjacent frames. As a result,
continuous clicking sounds are clearly evident when playing the watermarked audio
signal. To get rid of such noise, Hanning windowing and half overlapping (i.e.,
overlap factor p D 12 ) are utilized to smooth the edges.
Secondly, since SMM is a statistical method, repetition coding is expected to
help increase the overall detection accuracy. This is because repetition coding can
help maintain the statistical properties of successive frames. Therefore, nr times
repetition coding (nr D 3 or 5) is employed in the simulations.
Thirdly, detection threshold Td is estimated by performing a number of pre-
attack experiments as follows. The watermarked signal is pre-attacked by some
Algorithm 3.2 Pseudocode of cepstrum domain watermarking
% Watermark embedding
% Note: Host signal so .n/ ; 1 n No is split into NP frames fgi g with N samples,
˘
% where NP D NNo . Also, Nw D NP .
sw D Œ ; % Initialize the watermarked signal sw .n/

so .n/ D fgi .j /g, 1 i NP and 1 j N ;
Binary watermark wo .i /, 1 i Nw ;
% Loop through every frame gi

for i D 1 W NP
% Function cceps ./ is to compute the real part of the complex cepstrum.
ci D cceps .gi /;
D mean .ci /
i
if wo .i / D 1, then ciw D ci i C ˛w ;
if wo .i / D 0, then ciw D ci i ˛w ;
% Function i cceps ./ is to compute the inverse complex cepstrum.
giw D i cceps ciw ;

sw D sw ; giw
end
% Watermark detection
% Notes: sa .n/ is the attacked signal.
˚
sa .n/ D gia .j / , 1 i NP and 1 j N ;
Predefined threshold Td ;
% Loop through every frame gia

for i D 1 W NP
cia D cceps gia ;
i D sum cia ;
if i Td , then we .i / D 1;
if i < Td , then we .i / D 0.
end
Bit "1"
No attack: T d =3.2716 Echo addition: T d =3.6284
Bit "0"
4 4
2 2
μi
μi
0 0
−2 −2
−4 −4
1 100 200 249 1 100 200 249
Frame index Frame index

MP3 compression (48kps): T d =3.1196 Cropping: T d =1.0419
4 4
2 2
μi
μi
0 0
−2 −2
−4 −4
1 100 200 249 1 100 200 249
Frame index Frame index
Fig. 3.7 Distributions of Rone and Rzero under different attacks. Note that these data are produced
by watermarking with N D 4; 096, ˛w D 0:001, and nr D 3
commonly used attacks, such as noise addition, low-pass filtering, echo addition,
MP3 compression, and random samples cropping. For every attacked signal, the
sum of the cepstrum coefficients of each frame is written as i D sum cia .
If bit “1” was originally embedded in the ith frame, i is put into the “one”
set, i.e., Rone D f i j wo .i / D 1g. Otherwise, i is put into the “zero” set, i.e.,
Rzero D f i j wo .i / D 0g. Then, the minimum of Rone and the maximum of Rzero
0
are averaged to be one sub-threshold Td D min.Rone /Cmax.R
2
zero /
. Experimental results
show that the elements in Rone are positive numbers, commonly being around N˛w ,
whereas the elements in Rzero vary greatly between ŒN˛w ; N˛w . Therefore, the
maximum sub-threshold is a proper Td to achieve low BERs. To further reduce
the BER, more delicate adjustments on Td are required. For a better illustration,
the distribution of elements in Rone and Rzero is plotted in Fig. 3.7, where they
are denoted by blue asterisks and red circles respectively. As shown on the graph,
0
the maximum sub-threshold Td D 3:6284 under echo addition is an appropriate
Td . Furthermore, after subtle alteration, Td D 3:5 is utilized as the final detection
threshold for the conditions of N D 4; 096, ˛w D 0:001, and nr D 3 in Table 3.4.
Finally, variable frame length in the detection is utilized to combat pitch-invariant
time-scale modification (PITSM). Under the default setting for SMM detection, the
attacked signal sa .n/ ; 1 n Na is assumed to be as long as the watermarked
Table 3.4 Results of performance evaluation of cepstrum domain watermarking

Watermarking strength ˛w 0.001 0.0015 0.001
Watermark length Nw 504 168 100 504 168 100 251 83 50
Repetition coding nr 1 3 5 1 3 5 1 3 5
Detection threshold Td 1:6 2:5 3:5
SNR /dB 27:07 26:49 26:73 20:74 20:28 20:06 17:92 17:37 17:37
ODG 0:16 0:17 0:16 0:46 0:44 0:49 0:04 0:08 0:07
No attack 25:99 8:93 2:00 18:45 0 0 11:55 0 0

Noise (30 dB) 27:98 10:12 3:00 18:25 0 0 11:55 0 0
Resampling (22.05 kHz) 26:79 8:33 3:00 18:45 0 0 11:55 0 0
Requantization (8 bit) 26:59 9:52 2:00 18:25 0 0 11:55 0 0
Amplitude +10 % 27:18 8:93 4:00 18:45 0 0 11:55 0 0
-10 % 27:78 10:12 5:00 19:44 0 0 12:35 0 0
Lp filtering (8 kHz) 26:39 8:33 2:00 18:25 0 0 13:15 0 0
DA/AD (line-in jack) 48:41 55:36 52:00 48:41 51:19 53:00 53:78 57:83 56:00
Echo (0.3, 200 ms) 27:58 2:38 4:00 20:83 0:60 0 13:94 0 0
Reverb (1 s) 28:17 12:50 4:00 18:85 0 0 14:34 0 0
(continued)
71
72
Table 3.4 (continued)

Watermarking strength ˛w 0.001 0.0015 0.001
Compression I 96 kbps 25:99 8:33 3:00 18:45 0 0 12:75 0 0
64 kbps 26:79 10:71 3:00 17:86 0 0 12:35 0 0
48 kbps 27:58 10:71 3:00 19:05 0 0 11:55 0 0
Cropping (4 25 ms) 45:83 35:71 22:00 42:66 29:76 24:00 35:46 16:87 0
Jittering (0.1/20 ms) 44:25 31:55 11:00 38:10 17:26 1:00 28:69 0 0
Inserting (4 25 ms) 49:21 39:29 24:00 42:06 29:76 23:00 31:47 18:07 0
PITSM C10 % N 49:40 45:83 45:00 46:43 47:02 46:00 52:19 48:19 44:00
NQ 33:73 9:52 8:00 26:79 0:60 0 17:93 0 0
10 % N 51:18 48:39 49:46 49:03 49:03 41:94 48:71 48:05 36:96
NQ 34:72 16:07 13:00 24:6 0:60 0 15:14 0 0
TPPSM C10 % 32:94 8:33 5:00 26:98 1:19 0 17:53 0 0
10 % 33:53 11:90 10:00 23:41 0:60 0 10:76 0 0
3 Audio Watermarking Techniques
signal (as well as the host signal), i.e., Na D No . Also, the detection uses the same
frame length as the embedding, i.e., NQ D N . Then, the same splitting j method
k
Q
adopted in the embedding is employed to divide sa .n/ into NQ w D N Q
a p N
D
j k N .1p/
No pN
N .1p/
D Nw frames. Correspondingly, Nw watermark bits are extracted.
However, PITSM will adjust playback speed of the audio signal and Na is
changed accordingly. For example, it is lengthened by positive PITSM .Na > No / or
shortened by negative PITSM .Na < No /. For most audio watermarking techniques,
such alteration in signal length will cause severe problems to the detection. On one
hand, both positive and negative PITSM modify the time-scale of the watermarked
signal, which results in a displacement between the detection and embedding.
Without the retrieval of synchronization, watermark detection cannot work properly.
On the other hand, in the case of negative PITSM, we probably are unable to extract
as many watermark bits as are embedded. This is because the attacked signal has not
enough samples for the detection. For example, given No D 5:252 105 and N D
j k
No pN 5:252105 12 2;048
2; 048, the host signal is split into Nw D N .1p/ D 2;048.1 12 /
D 511
frames, and hence a number of bits up to 511 can be embedded. After being modified
by 10 % PITSM, the attacked signal has a shorter length of Na D 4:77455 j k105 .
Q
If NQ D N D 2; 048 is still in use, there would be NQ w D N a p N
NQ .1p/
D

1
4:7745510 2 2;048
5
2;048.1 12 /
D 465 frames. This means that only 465 bits can be extracted
8
at most, not to mention the low detection accuracy.
Under such circumstances, variable frame length for re-synchronization in the
Q
detection isjproposed: k the frame length N should vary with signal length Na ,
Q
i.e., N D N so . So the attacked signal would still be split into NP frames
sa
approximately.
MATLAB script for cepstrum domain watermarking can be found as
C epst rum wat ermarki ng:m file under Audio_Watermarking_Techniques
folder in the attached CD.
It is worth mentioning that although other desynchronization attacks such as
random samples cropping, zeros inserting, and jittering may also change signal
length, variable frame length is not very effective in these cases, especially the
former two. Under serious cropping and inserting attacks, a large amount of
removed or added samples occur locally, not uniformly along the whole water-
marked signal. Therefore, the detection with variable frame length cannot achieve
proper re-synchronization to recover the watermark. To withstand these attacks,
either the watermark is embedded on the basis of attack-invariant features, such
as the statistical mean of the cepstrum coefficients on a large scale, or the detection
can locate the positions where the attacks take place, such as the synchronization
method introduced in the next chapter.
8
Despite the fact that only a portion of bits are extracted, the corresponding BER is always
calculated for performance evaluations.
a
0.7
−0.7
0 2 4 6 8 10
b
0.7
−0.7
0 2 4 6 8 10
c
0.03
−0.03
0 2 4 6 8 10
Time (s)
Fig. 3.8 Host signal and a watermarked signal by cepstrum domain watermarking. Note that the
watermarked signal is produced by watermarking with N D 2; 048, ˛w D 0:0015 and nr D 3.
(a) Host audio signal. (b) Watermarked audio signal. (c) Difference between the watermarked and
host audio signals
The watermarked signal illustrated in Fig. 3.8 is generated by improved cepstrum

domain watermarking. Also, Table 3.4 shows the results of performance evaluation.
Considering the trade-off between imperceptibility and robustness, watermarking
strength ˛ is set to be 0.001 or 0.0015. One point to note is that cropping and jittering
would shorten the watermarked signal as well. In view of this, we usually do not
embed watermark bits at full capacity, but slightly less bits. As calculated above,
if N D 2048, the maximum capacity is Nw D 511 bits. In the experiments with
nr D 1, however, only 504 bits are embedded.
From Table 3.4, the overall performance of improved cepstrum domain water-
marking is attractive. On one hand, the watermarked signals have pretty good
perceptual quality in terms of high ODG scores. Informal subjective listening tests
also show that the watermarked signals are perceptually undistinguished from the
host signal.
On the other hand, with the help of repetition coding, the watermarked signals are
quite robust against all the attacks except for DA/AD conversion. Since each frame’s
statistical property has been changed by half overlapping between neighboring
frames, repetition coding is indispensable to watermark detection. For example, in

the case of N D 2; 048 and ˛w D 0:001, the watermarks without repetition coding
(nr D 1) have large BERs (more than 25 %). After triple repetition coding (nr D 3),
all the BERs excluding those under DA/AD conversion and some desynchronization
attacks are dropped to 10 % or so. The BERs can be further reduced by quintuple
repetition coding (nr D 5). Also, a longer frame length would help increase the
robustness as well. With the same watermarking strength and repetition coding,
the BERs of the watermarks with N D 4; 096 are much lower than those of
N D 2; 048. In addition, repetition coding and longer frame play an important role
in improving the robustness against cropping and inserting, whereas watermarking
strength is not really effective. For example, when N D 4; 096 and nr D 5, the
watermarks attacked by cropping and inserting can be detected without any errors.
This is because the statistical mean would keep stable on a large scale, which is
required for correct detection of the watermark. However, loss of capacity is the
price to pay for the robustness achieved by employing repetition coding and a longer
frame.
Experimental results show that the detection with variable frame length (indicated
as NQ ) can successfully extract the watermarks from the watermarked signals
attacked by PITSM. Moreover, the embedded watermarks are fairly resistant to
TPPSM, where the BERs under ˙10 % TPPSM are still very low. This indicates
that the statistical mean is immune to frequency variation caused by TPPSM.
On the whole, the improved cepstrum domain watermarking performs well
in terms of imperceptibility and robustness. However, the issue of security is
a challenge in real applications. Without prior knowledge, the attacker could
deliberately make a slight modification to the mean of cepstrum coefficients of the
watermarked signal, so that the watermark cannot be extracted properly.
3.2.5 Wavelet Domain Watermarking
In wavelet domain watermarking, the watermark is embedded into wavelet coef-

ficients of host audio signal. Compared to Fourier transform, wavelet transform is
more suitable to generate the time–frequency representation of nonstationary signals
such as audio signals [100].
As shown in Fig. 3.9, DWT decomposition and reconstruction involve the
multiresolution analysis and synthesis. During the decomposition, the input signal
is first decomposed into two parts: the high-frequency part (i.e., the “detail”
coefficients) from high-pass filter .H0 / and the low-frequency part (i.e., the
“approximation” coefficients) from low-pass filter .G0 /. Then, the low-frequency
part is further decomposed into high- and low-frequency parts. This process is
repeated until the desired level is reached. At each decomposition level, the time
resolution is halved and the frequency resolution is doubled. In this way, DWT
brings good frequency resolution at low frequencies and good time resolution
at high frequencies. Since the human ear is more sensitive to low-frequency
a
H0 2 cD1
S0
H0 2 cD2
G0 2
H0 2 cD3
G0 2
G0 2 cA3
b cD1 2 H1
S0
cD2 2 H1
2 G1
cD3 2 H1
2 G1
cA3 2 G1
Fig. 3.9 A three-level DWT decomposition and reconstruction. (a) Wavelet decomposition.
(b) Wavelet reconstruction.
Notes: 1. and denote downsampling and upsampling by two, respectively.
2. H0 =G0 and H1 =G1 are high-pass/low-pass analysis and synthesis filters, respectively
sounds, DWT exhibits similar characteristics in time–frequency resolution as the

human ear [101]. As for the reconstruction, the original signal can be perfectly
reconstructed from the DWT coefficients when analysis and synthesis filters satisfy
the following conditions [102]:
H0 .z/ G0 .z/ C H1 .z/ G1 .z/ D 0;

(3.13)
H0 .z/ G0 .z/ C H1 .z/ G1 .z/ D 2:
Moreover, if both high- and low-frequency parts are iteratively decomposed,
DWT is evolved into discrete wavelet packet transform (DWPT), which offers a
more complex and flexible analysis of audio signals. For example, DWPT-based
decomposition is employed for psychoacoustic modelling in [103].
3.2.5.1 Algorithm
Considering the good time–frequency resolution property of DWT, some audio

watermarking schemes in the wavelet domain are proposed, such as [7, 101, 104–
107]. Both methods in [101, 104] belong to wavelet domain watermarking using the
SS method. In [104], a perceptually shaped PRS was taken as the watermark and
spread over the detail coefficients at each level. To keep the imperceptibility and
robustness, Hwang et al. [101] embedded the watermark into post-masking regions
with high-energy and low zero-crossing rate (ZCR). Instead of DWT coefficients,
certain detail coefficients of the last level from the DWPT were chosen for
watermarking in [105]. Different from [7, 101, 104–106] embedded the watermark
in wavelet domain using the patchwork method. Moreover, to enhance the security,
Cvejic et al. [7] employed a secret key to randomly select the subbands used for
watermarking.
Inspired by the idea of cepstrum domain watermarking in [95], Li et al. [107]
applied SMM in the wavelet domain, where the mean of approximation coefficients
at the last level is modified to embed the watermark. Basically, the procedures of
watermark embedding and detection described in [107] are similar to Algorithm 3.2.
To further improve the performance, we also employ the Hanning window and
half overlapping for smooth transition as well as variable frame length to strive
against PITSM. However, the estimation of detection threshold Td is not necessary
in wavelet domain watermarking based on SMM. Similar pre-attack experiments as
those in cepstrum domain watermarking show that the detection threshold Td can
be specified as Td D 0.
MATLAB script for wavelet domain watermarking can be found as W avelet
wat ermarki ng:m file under Audio_Watermarking_Techniques folder in the
attached CD.
Figure 3.10 shows an example of the watermarked signal by wavelet domain

watermarking based on SMM. In our implementation, a three-level DWT decom-
position with “db4” wavelet [107] is carried out. Accordingly, the mean of DWT
coefficients cA3 is modified to embed the watermark. Considering the trade-off
between imperceptibility and robustness, the watermarking strength ˛w is set to be
0.01 or 0.02.
Comparing Table 3.5 with Table 3.4, wavelet domain watermarking based on
SMM performs slightly better than cepstrum domain watermarking. However,
DA/AD conversion is still a disaster to the survival of watermarks. As indicated by
high ODG scores, the watermarked audio signals are perceptually undistinguished
from the host signal, and informal subjective listening tests have also proved this.
From Table 3.5, one point to note is that high SNRs are not equivalent to high
ODG scores. The watermarked signals under the condition of N D 1; 024 and
˛w D 0:02 have lower SNRs, but higher ODG scores than those of N D 2; 048 and
˛w D 0:01. Seeing that the SNR is not a reliable indicator of perceptual quality for
audio watermarking, more objective quality measures are investigated in Chap. 6.
Similarly, repetition coding can greatly improve the detection of wavelet domain
watermarking based on SMM. When ˛w D 0:01 and N D 1; 024, the watermarks
generated with triple repetition coding are already able to combat various attacks
from noise addition to Compression I at different bitrates, excluding DA/AD
conversion. The resulting BERs are less than 5 %. Moreover, by using N D 2; 048
and variable frame length (indicated as NQ ), the watermarked signals are quite robust
a
0.7
−0.7
0 2 4 6 8 10
b
0.7
−0.7
0 2 4 6 8 10
c
0.02
−0.02
0 2 4 6 8 10
Time (s)
Fig. 3.10 Host signal and a watermarked signal by wavelet domain watermarking. Note that the
watermarked signal is produced by watermarking with N D 2; 048, nr D 3, and ˛w D 0:01. (a)
Host audio signal. (b) Watermarked audio signal. (c) Difference between the watermarked and host
audio signals
against all the attacks except DA/AD conversion, cropping, and inserting. To further
reduce the BERs under cropping and inserting, a longer frame (e.g., N D 4; 096)
would be required.
As a kind of statistical watermarking, wavelet domain watermarking based
on SMM is also troubled with security. In [95], the watermark was encrypted
before embedding for the purpose of security consideration. In this way, the
encrypted watermark remains incomprehensible to the attacker without the secret
key for decryption. However, the encryption only offers additional security on top
of watermarking, but cannot prevent deliberate alteration on the mean of DWT
coefficients from seriously destroying the embedded watermark. Therefore, it is
quite important to enhance the security of the watermarking scheme.
3.2.6 Echo Hiding
Echo hiding embeds the watermark into host signals by introducing different echoes.
With well-designed amplitudes and delays (offset), the echoes are perceived as
resonance to host audio signals and would not produce uncomfortable noises [36].
Table 3.5 Results of performance evaluation of wavelet domain watermarking
Embedding strength ˛w 0.01 0.02 0.01
Watermark length Nw 1; 016 338 203 1; 016 338 203 504 168 100

SNR /dB 23:38 22:92 22:90 20:80 20:01 20:06 26:28 25:90 25:59
ODG 0:15 0:15 0:16 0:10 0:13 0:12 0:48 0:47 0:50
No attack 22:34 2:07 0 16:34 0:30 0 18:45 0 0

Noise (3 dB) 22:74 2:37 0 16:34 0:59 0 19:05 0 0
Resampling (22.05 kHz) 22:83 2:37 0 16:93 0:30 0 18:25 0 0
Requantization (8 bit) 22:44 2:07 0 16:14 0:30 0 19:25 0 0
Amplitude C10 % 22:34 2:07 0 16:34 0:30 0 18:45 0 0
10 % 22:34 2:07 0 16:34 0:30 0 18:45 0 0
Lp filtering (8 kHz) 21:75 2:07 0 16:14 0:30 0 18:65 0 0
DA/AD (line-in jack) 47:24 47:04 50:74 47:83 47:93 46:80 48:41 48:21 55:91
Echo (0.3, 200 ms) 25:79 4:44 3:45 19:69 1:18 0 21:63 1:79 0
Reverb (1 s) 23:33 4:44 0:99 15:26 0:30 0 19:25 1:79 0
(continued)
79
80
Embedding strength ˛w 0.01 0.02 0.01
Compression I 96 kbps 22:05 2:66 0:49 17:03 0:30 0 19:05 0:60 0
64 kbps 22:34 2:66 0 17:03 0:30 0 18:65 0:60 0
48 kbps 21:95 2:96 0:49 16:83 0:59 0 18:85 0 0
Cropping (4 25 ms) 48:03 38:76 35:96 44:19 39:94 33:50 37:70 32:14 19:00
Jittering (0.1/20 ms) 46:85 41:72 33:00 44:49 31:36 21:67 38:69 22:02 2:00
Inserting (4 25 ms) 46:36 42:90 36:45 42:62 41:42 33:99 40:67 35:71 18:00
PITSM C10 % N 48:62 47:04 44:33 49:70 48:82 46:31 50:40 54:76 46:00
NQ 37:30 24:26 12:81 37:11 13:31 6:40 28:77 4:17 0
10 % N 49:52 49:03 46:77 51:45 54:52 42:47 52:04 47:10 45:16
NQ 38:29 21:30 7:88 34:25 9:17 2:96 30:95 6:55 1:00
TPPSM C10 % 37:40 18:05 7:88 35:04 12:43 2:96 25:99 8:33 2:00
10 % 32:87 14:50 6:90 29:63 4:44 1:97 26:79 4:17 1:00
a b
Host signal Host signal
Echo for bit “1” Echo for bit “0”
a a
d1 Time d0 Time
Fig. 3.11 Impulse response of echo kernels. (a) “One” kernel. (b) “Zero” kernel
3.2.6.1 Algorithm
In the embedding process, the watermarked signal sw .n/ is generated by the

convolution between host signal so .n/ and echo kernel h .n/. The basic echo hiding
scheme employs a single echo kernel, whose impulse response is expressed as
h .n/ D ı .n/ C ˛ı .n d / ; (3.14)
where ˛ is echo amplitude and d is the delay. To represent bit “1” and “0,” echo
kernels are created with different delays (d1 and d0 ), as shown in Fig. 3.11. Usually,
the allowable delay offsets for 44.1 kHz sampled audio signals are set to be 100
150 samples (about 2:3 3:4 ms) [108]. Consequently, the watermarked signal is
described as follows:
sw .n/ D so .n/ ˝ h .n/

D so .n/ C ˛ so .n d / : (3.15)
In order to detect the watermark, cepstrum analysis is utilized to discern the value
of delay. According to Fig. 3.6a, the complex cepstrum of the watermarked signal
sOw .n/ is defined as
sOw .n/ D F1 flog . F fsw .n/g /g ; (3.16)
where F fg and F1 fg denote the Fourier transform and the inverse Fourier trans-
form, respectively. After substituting Eq. (3.15) into Eq. (3.16), sOw .n/ is written as
sOw .n/ D F1 flog .F fso .n/ ˝ h .n/g/g

D F1 flog .F fso .n/g/g C F1 flog .F fh .n/g/g
˚ ˚
D F1 log So e j! C F1 log H e j!
D sOo .n/ C hO .n/ ; (3.17)

˚ ˚
where sOo .n/ D F1 log So e j! and hO .n/ D F1 log H e j! are the
complex cepstrum of so .n/ and h .n/, respectively. In view of Eq. (3.14), we have
2 3
H e j! D 1 C ˛e j!d . Using the Taylor series log .1 C x/ D x x2 C x3
for jxj < 1, hO .n/ is calculated by
˚
hO .n/ D F1 log 1 C ˛e j!d

˛2 ˛3
D F1 ˛e j!d e j 2!d C e j 3!d
2 3
˛2 ˛3
D ˛ı .n d / ı .n 2d / C ı .n 3d / (3.18)
2 3
Accordingly, sOw .n/ in Eq. (3.17) becomes
˛2 ˛3
sOw .n/ D sOo .n/ C ˛ı .n d / ı .n 2d / C ı .n 3d / (3.19)
2 3
This shows that a series of impulses with exponentially decaying amplitudes
repeatedly appear for every d samples. In particular, the dominant spike is just
located at the delay (n D d ) and its amplitude is equal to that of the embedded
echo, ˛. Then, the watermark can be decided based on the comparison between the
values of cepstrum coefficients at two delays, i.e., sOw .d1 / and sOw .d0 / .
To further increase the amplitude of cepstrum spikes representing the echoes,
autocorrelation of the cepstrum (auto-cepstrum) ca .n/ is employed to detect the
delay [36, 109]:
n o
ca .n/ D F1 log .F fsw .n/g/2 : (3.20)
Since the autocorrelation calculates the signal power at each delay, the power
spike in the cepstrum is more prominent, as illustrated in Fig. 3.12. Therefore, the
watermark bit is determined by comparing ca .d1 / and ca .d0 /:
8
<1; if ca .d1 / ca .d0 / ;
we .i / D (3.21)
:0; if ca .d1 / < ca .d0 / :
The performance of echo hiding depends on echo kernels, and hence different
echo kernels are introduced to improve the imperceptibility and robustness of the
embedded echoes [3, 108, 110–112]. In [108], the echo kernel comprises multiple
positive and negative echoes with different delays. Typically, a dual echo kernel
with one positive and one negative echo is denoted as
h .n/ D ı .n/ C ˛1 ı .n d / ˛2 ı .n d 4/ ; (3.22)

10
Complex cepstrum
8 Real cepstrum
Auto−cepstrum
6
2
Amplitude
−2 Close−up
1
−4
0.5
−6
−8 0
401
−10
1 401 1024 2048 3072 4096
Samples
Fig. 3.12 Auto-cepstrum for echo detection
where j4j 5 samples. Then, its cepstrum hO .n/ is calculated as

˛ 2 C ˛22
hO .n/ D ˛1 ı .n d / ˛2 ı .n d 4/ C 1 fı .n 2d / C ı .n 2d 24/g
2
C ˛1 ˛2 ı .n 2d 4/ C (3.23)
Therefore, the watermark bit can be determined by comparing Œca .d1 / ca .d1 C4/
and Œca .d0 / ca .d0 C 4/. By virtue of the positive and negative echo kernel,
high-energy echoes can be added to enhance the robustness, while audio quality
is not deteriorated. This is because by combining these closely located positive and
negative echoes, the frequency response of the dual echo kernel can remain flat
over lower frequencies. Thus, the perceptual quality of the watermarked signal is
preserved.
Later, Kim et al. [110] proposed the backward and forward echo kernel as
follows:
h .n/ D ı .n/ C ˛ı .n d / C ˛ı .n C d / : (3.24)
Then, its cepstrum hO .n/ is calculated as
hO .n/ D ˛ fı .n d / C ı .n C d /g ˛2
2 fı .n 2d / C 2ı .n/ C ı .n C 2d /g
˛3
C 3 fı .n 3d / C 3ı .n d / C 3ı .n C d / C ı .n C 3d /g
D ˛ C ˛ 3 C ˛ 5 C ı .n d / C
D ˛ 2 ı .n d / C :
1˛
(3.25)
Therefore, the watermark bit can be determined by comparing ca .d1 / and ca .d0 /.
From Eq. (3.25), the amplitude of the cepstrum peak at n D d is equal to 1˛ ˛
2,
which is larger than ˛ for 0 < ˛ < 1. As a result, the detection rate is increased.
MATLAB script for echo hiding can be found as Echo hid i ng:m file under
Audio_Water-marking_Techniques folder in the attached CD.
In addition, a time-spread echo kernel [111] is introduced to enhance the security.
Although the large amplitude of the cepstrum peak is beneficial to robustness,
obvious spikes are against the purpose of security and unauthorized attackers might
detect the existence of echoes easily without prior knowledge. By using a PRS to
spread multiple echoes, the amplitude of each echo becomes small. It contributes
directly to the imperceptibility, while the detection ability is better maintained as
well. Furthermore, log-scaling watermark detection [112] is proposed to cope with
pitch-scale modification (PSM). Recently, Chen et al. [3] designed an advanced echo
hiding scheme based on the analysis-by-synthesis approach.
In our experiments, echo hiding schemes with a single echo kernel (Kernel 1), a
positive and negative echo kernel (Kernel 2), and a backward and forward echo
kernel (Kernel 3) are evaluated. For a fair comparison, the frame length and the
amplitude of different echo kernels are the same, i.e., N D 4; 096 and ˛ D 0:2.
In order to enhance the security, a sequence of pseudorandom numbers is utilized
as the secret key to shift between several echo delays. Each delay is denoted as
dxy , where x and y represents the pseudorandom number (PRN) and the watermark
bit, respectively. In this way, if the PRN is 1 and bit “0” is to be embedded, then
d10 is selected. In considering both imperceptibility and robustness, the value of
delays is set as follows: d11 D 100 and d01 D 120 are used for embedding bit
“1,” while d10 D 110 and d00 D 130 are used for embedding bit “0.” Moreover,
additional delay 4 D 4 is used in the positive and negative echo kernel. Similar
to previous watermarking techniques, Hanning windowing and half overlapping for
smooth transition, variable frame length to combat PITSM and repetition coding are
also employed to further improve the performance.
Figure 3.13 shows the watermarked signal produced by echo hiding using the
positive and negative echo kernel. The results of performance evaluation of three
echo kernels are summarized in Table 3.6. The positive and negative echo kernel
provides higher SNRs than the other two kernels, which means the least distortion
between the watermarked and host signals. However, the watermarked signals with
three echo kernels obtain the same ODG scores and are deemed to be similar in
perceptual quality. Informal subjective listening tests show that the added echoes do
not introduce annoying noises, rather they make the sound rich.
Regarding the robustness of the watermarked signals, the backward and forward
echo kernel (Kernel 3) generally provides the best detection rate. With the help
of triple repetition coding (nr D 3) and variable frame length (indicated as NQ ),
the BERs of watermarks under all attacks except TPPSM are less than 10 %. After
a
0.7
−0.7
0 2 4 6 8 10
b
0.7
−0.7
0 2 4 6 8 10
c
0.1
−0.1
0 2 4 6 8 10
Time (s)
Fig. 3.13 Host signal and a watermarked signal by echo hiding. Note that the watermarked signal
is produced by watermarking with N D 4; 096, ˛ D 0:2, and nr D 3. (a) Host audio signal.
(b) Watermarked audio signal. (c) Difference between the watermarked and host audio signals
quintuple repetition coding (nr D 5), the BERs under cropping and inserting are
reduced further. For the other two kernels, the positive and negative echo kernel
(Kernel 2) performs slightly better than the single echo kernel (Kernel 1), where the
BERs are decreased on average. In addition, it is worth mentioning that all three
kernels exhibit good resistance to DA/AD conversion. However, all the watermark
detections with any kernel fail completely under TPPSM. Also, echo addition might
be a hazardous attack on echo hiding watermarking. If echo delays in the attack
happen to be the same as those of echo kernels, the mistakes in watermark detection
are unavoidable.
Echo hiding is a watermarking technique specifically for audio signals. By
selecting the proper amplitude and delay of echo kernels, the echoes embedded as
the watermark can be imperceptible and robust against most attacks. However, echo
hiding may suffer from two deficiencies. One is weak security, because obvious
cepstrum peaks might be tampered with deliberately. The other is about inborn
echoes contained in natural sound, which might result in false-positive errors [110].
86
Table 3.6 Results of performance evaluation of echo hiding
Echo hiding method Kernel 1 Kernel 2 Kernel 3
Watermark length Nw 251 83 50 251 83 50 251 83 50
SNR /dB 14:78 14:19 14:13 19:25 18:48 18:30 11:56 11:45 10:87
ODG 2:20 2:20 2:20 2:20 2:20 2:20 2:20 2:20 2:20
No attack 27:09 4:82 2:00 22:71 2:41 0 12:75 0 0

Noise (30 dB) 35:86 21:69 14:00 32:27 20:48 12:00 23:51 6:02 2:00
Resampling (22.05 kHz) 24:70 9:64 2:00 27:49 6:02 2:00 17:13 1:20 0
Requantization (8 bit) 33:47 18:07 16:00 35:46 14:46 14:00 23:90 6:02 4:00
Amplitude C10 % 27:09 3:61 2:00 22:31 3:61 0 13:94 0 0
10 % 27:89 3:61 2:00 23:90 2:41 2:00 14:34 0 0
Lp filtering (8 kHz) 28:29 9:64 2:00 25:90 8:43 4:00 15:54 1:20 0
DA/AD (line-in jack) 23:11 12:05 4:00 20:72 4:82 2:00 12:35 2:41 0
Echo (0.3, 200 ms) 28:69 7:23 8:00 21:12 8:43 4:00 16:33 1:20 2:00
Reverb (1 s) 24:70 6:02 2:00 21:51 4:82 0 15:14 0 0
Compression I 96 kbps 26:69 6:02 6:00 24:30 3:61 0 17:53 1:20 0
64 kbps 23:51 10:84 2:00 20:32 3:61 0 17:13 1:20 0
48 kbps 29:88 9:64 2:00 23:90 8:43 0 17:13 2:41 0
Cropping (4 25 ms) 41:83 19:28 6:00 38:25 14:46 6:00 32:67 9:64 2:00
Jittering (0.1/20 ms) 37:45 16:87 10:00 31:87 10:84 4:00 25:10 3:61 0
Inserting (4 25 ms) 39:04 15:66 4:00 35:46 18:07 4:00 28:69 8:43 2:00
PITSM C4 % N 47:81 43:37 28:00 47:81 34:94 24:00 42:23 37:35 36:00
NQ 18:73 6:02 2:00 22:71 6:02 0:00 16:33 2:41 0
4 % N 45:69 38:96 21:74 44:83 36:36 23:91 47:84 45:45 23:91
NQ 23:90 8:43 0:00 24:30 2:41 2:00 15:54 4:82 2:00
TPPSM C4 % 44:22 38:55 40:00 43:82 50:60 34:00 45:02 32:53 36:00
4 % 46:22 42:17 48:00 59:36 80:72 80:00 46:22 43:37 50:00
87
3.2.7 Histogram-Based Watermarking
As opposed to the previous methods, histogram-based watermarking works at

the global rather than the local characteristic of host signals, i.e., modifying the
histogram to embed the watermark. Histograms are well suited to describe the
distribution of large data sets. In digital watermarking, the histogram is obtained
by dividing a range of sample values into equal-sized bins and then calculating the
number of samples occurring in each bin. The samples used might be from the entire
or only a part of the host signal.
3.2.7.1 Algorithm
There are several ways in which the histogram is modified to embed the watermark.
Coltuc et al. [113] implemented a robust image watermarking using exact histogram
specification, where the image histogram is shape-altered (e.g., a saw-teeth shape
with 3 Ï 8 periods) to represent a watermark. Later, MeÅŸe et al. [114] designed
an optimal algorithm for histogram modification, so that the mean square error
(MSE) is minimized between the modified and host images. In [115], the watermark
was embedded into the image by permuting some pairs of histogram bins. Apart
from being robust against geometrical attacks, the watermarking scheme is also
reversible, which means that the watermarked image can be fully restored to its
original status. Unlike the commonly used histogram in the time domain, Xuan
et al. [116] employed histogram shifting in the integer wavelet transform domain
for reversible image watermarking.
In [35], histogram modification is applied in audio watermarking. Based upon
the fact that the modified audio mean and the audio histogram shape are invariant
to temporal scaling, the authors designed an audio watermarking scheme resistant
to time-scale modification (TSM), random cropping, and inserting attacks. The
modified audio mean A is defined as the average of the absolute value of the
XNo
16-bit signed audio signal, i.e., A D N1o jso .n/j. In the embedding, A is used
nD1
the amplitude range B of the samples for producing the histogram, i.e.,
to decide
B D A; A . Through extensive experiments on different audio signals, a
suggested range is 2 Œ2; 2:5[35]. The number of histogram bins Nbin depends
on watermark length Nw , i.e., Nbin 3Nw . The factor 3 comes from the reason
that the watermark is embedded by controlling the relative relation of the number
of samples in every threej neighboring
k bins. Accordingly, the bin width Nbw is
calculated as Nbw D N 2A
bin
. The value of Nbw affects the properties of both
imperceptibility and robustness. A small Nbw is likely to reserve the shape of the
original histogram, which is beneficial to the perceptual quality. Meanwhile, each
bin should contain sufficient samples to ensure the watermark robustness. Therefore,
must be carefully chosen to obtain a suitable Nbw .
After the histogram is constructed, its bins are divided into Nw groups, each of
which has three bins. For every group, the number of samples in three consecutive
bins is denoted as Nb1 , Nb2 , and Nb3 , respectively. Then, one watermark bit is
embedded into one group of histogram bins by applying the following rules [117]:
8
< .Nb1 CNb3 / E ; if wo .i / D 1;
2Nb2 h
(3.26)
: 2Nb2 E ; if w .i / D 0;
.Nb1 CNb3 / h o
where the embedding strength Eh is around 1:2 1:5. Obviously, a large Eh would
increase the robustness but degrade the perceptual quality.
From the embedding process, one point to note is that histogram-based water-
marking has the advantage of handling silent intervals of the host audio signal. As
mentioned in Sect. 3.1.1, it is nearly impossible to embed the watermark into zero
values. In histogram-based watermarking, all zero-value samples fall into the center
of the histogram. Therefore, these samples can be well preserved, provided that the
one or two bins right in the center are exempted from watermarking.
w
In the detection, the modified mean of the watermarked signal A is calculated,
and then the histogram is constructed with the same and Nbin . Furthermore,
the histogram bins are divided into groups in the same way as the embedding.
Therefore, the watermark bit is decided by comparing the number of samples in
w w w
three consecutive bins, namely, Nb1 , Nb2 , and Nb3 :
8
<1; .N w CN w /
if b12N w b3 1;
we .i / D b2 (3.27)
:0; .Nb1w CNb3w /
if 2N w < 1:
b2
w
Since the attacks might change the modified audio mean, A is probably not equal to
A. As a result, the histogram range is deviated and the watermark cannot be detected
correctly. Therefore, it is necessary to search for the proper modified audio mean.
Experimental results show that the fluctuation of the modified audio mean is usually
less than ˙6 %; thus, optimal searching within this range is proposed to prevent
exhaustive searching [35].
MATLAB script for histogram-based watermarking can be found as
H i st ogram wat ermarki ng:m file under Audio_Watermarking_Techniques
folder in the attached CD.
Figure 3.14 shows an example of the watermarked signal by histogram-based

watermarking[117]. The difference between the watermarked and host audio signals
is small, and accordingly the SNRs in Table 3.7 are fairly high. From Fig. 3.14c,
one point to note is that most samples in the front and rear part of the host signal
are changed to embed the watermark, but the middle part remains intact. This is
a
0.7
−0.7
0 2 4 6 8 10
b
0.7
−0.7
0 2 4 6 8 10
c
0.003
−0.003
0 2 4 6 8 10
Time (s)
Fig. 3.14 Host signal and a watermarked signal by histogram-based watermarking. Note that the
watermarked signal is produced by watermarking with Nw D 40, D 2:2, and Eh D 1:5.
(a) Host audio signal. (b) Watermarked audio signal. (c) Difference between the watermarked and
host audio signals
because the samples in each histogram bin are arranged in the order they are found
in the host signal. During the process of modifying the samples into three bins of
each group, the samples at the beginning and the end of each bin are always chosen
with priority. Our implementation attempted to randomly select the samples in the
bins, but the perceptual quality of the watermarked signal became worse. The reason
might be that the randomly distributed modification resembles the addition of white
noise.
The results of performance evaluation of histogram-based watermarking are
illustrated in Table 3.7. Generally, a larger Nw or a smaller would result in a
better perceptual quality, but a weaker robustness. This is because the bin width Nbw
is proportional to , but inversely proportional to Nw . Moreover, histogram-based
watermarking is indeed quite robust against most desynchronization attacks such
as TSM, cropping, jittering, and inserting. In compromise with the imperceptibility,
the watermarks are also able to survive from PSM to a certain extent.
However, the BERs of the watermarks are rather high in the cases of some attacks
like low-pass filtering, DA/AD conversion, reverberation, and MP3 compression.
w
Even extending the searching range of A to ˙10 %, the detection rate has no
substantial improvement.
Table 3.7 Results of performance evaluation of histogram-based watermarking
Watermark length Nw 20 40
Embedding range 2.2 2.5 2.2 2.5
Embedding strength Eh 1:2 1:5 1:2 1:5 1:2 1:5 1:2 1:5
SNR /dB 41:99 38:66 40:96 37:55 47:95 44:59 46:70 43:35
ODG 2:17 2:89 2:21 2:86 1:73 2:42 1:83 2:60
No attack 0 0 0 0 0 0 0 0
Noise (30 dB) 5:00 0 0 0 20:00 12:50 25:00 17:50
Resampling (22.05 kHz) 0 0 0 0 0 0 0 0
Requantization (8 bit) 30:00 30:00 35:00 35:00 35:00 35:00 27:50 25:00
Amplitude C10 % 0 0 0 0 0 0 0 0
10 % 0 0 0 0 0 0 0 0
Lp filtering (8 kHz) 5:00 5:00 5:00 0 27:50 22:50 22:50 17:50
DA/AD (line-in jack) 25:00 20:00 35:00 30:00 32:50 30:00 27:50 30:00
Echo (0.3, 200 ms) 25:00 30:00 30:00 30:00 15:00 22:50 20:00 22:50
Reverb (1 s) 25:00 25:00 15:00 15:00 27:50 22:50 17:50 20:00
(continued)
91
92
Watermark length Nw 20 40
Embedding range 2.2 2.5 2.2 2.5
Compression I 9 kbps 0 0 0 0 15:00 15:00 10:00 15:00
64 kbps 5:00 5:00 5:00 5:00 15:00 17:50 22:50 20:00
48 kbps 15:00 5:00 15:00 10:00 22:50 25:00 22:50 27:50
Cropping (4 25 ms) 0 0 0 0 0 0 0 0
Jittering (0.1/20 ms) 0 0 0 0 0 0 0 0
Inserting (4 25 ms) 0 0 0 0 0 0 0 0
PITSM C10 % 0 0 0 0 0 0 0 0
10 % 0 0 0 0 0 0 0 0
TPPSM C10 % 0 0 0 0 10:00 0 7:50 5:00
10 % 0 0 0 0 2:50 0 5:00 2:50
3.3 Summary 93
The reason is that the attacks have smoothed the histograms and the relative
w w w
relations between Nb1 , Nb2 , and Nb3 of each group are destroyed. As an exception,
the histogram changes dramatically after requantization. Since quantization rounds
off the samples, some histogram bins are eliminated, which leads to failure in
watermark detection. For an 8-bit quantizer, the quantization error is calculated as
o .n// o .n//
q D 2max.s
28
D max.s
128
[118]. Therefore, the width of histogram bins should be
larger than q to combat requantization attack.
Later, instead of the histogram in the time domain, Xiang et al. [2] exploited
the invariance of histogram shape in the low-frequency subband of the DWT
domain. It is reported that the improved watermarking method is more robust against
TSM, low-pass filtering, and MP3 compression. Nevertheless, the security of the
watermarking scheme is still a serious issue.
In addition to the mean, other statistical moments can be used as invariants in
histogram-based watermarking. For example, image steganalysis in [119] employed
the first four moments (i.e., the mean, variance, skewness, and kurtosis) of the
subbands that are decomposed by quadrature mirror filters (QMF). In the book, we
attempted to utilize the moments in [120] for audio watermarking, where the audio
signal has been converted into a two-dimensional square matrix as implemented in
[121]. However, most invariants (except for the first moments) are not applicable
to audio watermarking. The reason might be that the audio signal represented in a
two-dimensional form is incomparable to a two-dimensional image.
3.3 Summary
This chapter has mainly investigated different audio watermarking techniques,

such as LSB modification, phase coding, SS watermarking, cepstrum domain
watermarking, wavelet domain watermarking, echo hiding, and histogram-based
watermarking. We also contributed to the improvement of some of these techniques.
To achieve a better comprehension of various methods, specifications on the exe-
cution of performance evaluation were described. Firstly, seventeen pieces of audio
test signals in four categories were prepared for evaluation. The diversity of test
signals helps verify the applicability of audio watermarking techniques. Secondly,
the instructions for conducting perceptual quality assessment including subjective
listening tests and objective evaluation tests were presented. For subjective listening
tests, a group of trained participants are required to grade the watermarked signals
in MUSHRA test and the SDG rating. For objective evaluation tests, both the ODG
provided by software PEAQ and the SNR are adopted as impersonal measurements.
Thirdly, test items and their default parameters in basic and advanced robustness
tests were separately depicted in detail. A number of common signal operations,
desynchronization attacks, and advanced attacks are included in the robustness test.
Following that, different audio watermarking techniques were intensively studied.
LSB modification and phase coding are early audio watermarking techniques.
Although LSB modification might be unperceived, both methods performed

relatively poorly in the robustness test. SS watermarking is the most prevalent
technique for digital watermarking because of its robustness and security. However,
effective ways for SS watermarking to combat desynchronization attacks are
required. The implemented cepstrum domain watermarking and wavelet domain
watermarking were based on SMM. After our improvement, their performance
in terms of imperceptibility and robustness became fairly good. However, the
issue of security is a challenge for such statistical watermarking. Echo hiding is
a method especially for audio watermarking. With the proper echo kernel, the
echoes embedded as the watermark can be made imperceptible and resistant to most
attacks. Different from the previous techniques, histogram-based watermarking is
rather robust against desynchronization attacks, but somewhat vulnerable to several
common signal processing operations, such as low-pass filtering and reverberation.
Two observations can be obtained from the experiments. One is that trade-
offs always exist between imperceptibility, robustness, security, and data payload
in audio watermarking systems. The other is that there are no universal audio
watermarking techniques to combat all the attacks. In most cases, it is more difficult
to conquer desynchronization attacks.
Chapter 4
Proposed Audio Watermarking Scheme
Imperceptibility, robustness, and security are vital considerations in the design

of any audio watermarking scheme for copyrights protection. In this chapter, a
spread spectrum (SS)-based audio watermarking technique which involves the
psychoacoustic model, multiple scrambling, adaptive synchronization, frequency
alignment, and coded-image watermark is presented. To preserve the perceptual
quality of the watermarked signal, amplitude shaping using the psychoacoustic
model is employed. Also, the proposed scheme integrates multiple scrambling
operations into the embedding process to prevent unauthorized detection. That
is, the amount and position of the slots used for embedding each watermark bit
are randomly set and certain subbands are randomly selected for the embedding.
Moreover, adaptive synchronization and frequency alignment are developed to
retrieve the watermarks from the attacked watermarked signals that suffer loss of
synchronization. In addition, the information to be embedded can be encrypted with
a coded-image, so as to provide a semantic meaning for verification as well as extra
security.
The chapter is organized as follows. Section 4.1 describes the selection of
watermarking regions and the structure of the watermarking domain. In Sect. 4.2,
the embedding algorithm including multiple scrambling is presented. Section 4.3
focuses on watermark detection emerging with adaptive synchronization and fre-
quency alignment. This is followed by an introduction to the coded-image water-
mark in Sect. 4.4. Finally, Sect. 4.5 summarizes the characteristics of the proposed
audio watermarking scheme.
4.1 Preliminaries
Generally, the proposed audio watermarking scheme includes watermark embed-

ding and watermark detection. In watermark embedding, we embed not only
the watermark for copyrights protection, but also synchronization information for
watermark detection. These data are basically represented by the bits. In our scheme,
96 4 Proposed Audio Watermarking Scheme
the watermark bits are repeatedly embedded to increase the robustness. Also, a
synchronization bit is repeatedly embedded for synchronization purposes.
For a better understanding of the watermarking procedure, we start with some
preliminary knowledge, such as the selection of watermarking regions and the
structure of the watermarking domain.
4.1.1 Selection of Watermarking Regions
As mentioned in Sect. 3.1.1, embedding the watermarks into silent segments would
introduce unavoidable perceived noise. Therefore, a selection process is applied
on the host audio signal to determine the embedding regions. Correspondingly, it
is necessary to perform this procedure in the watermark detection stage, so as to
locate the regions for watermark detection. Since various attacks might alter the
watermarked signal, the watermarking regions should be rather stable to ensure that
they still can be identified in the attacked signals [4, 5, 122]. In this way, the process
of selection is a kind of initial synchronization between the watermark embedding
and detection.
There are several methods for selecting reliable watermarking regions which
usually follow certain distinct points, for example, salient point extraction [6,33,97],
peak point extraction [123], and envelope peak extraction [5]. Commonly, these
delicate methods have been employed as solutions to synchronization; however,
this is not necessary in the proposed scheme. The selection of our watermarking
regions mainly aims to preclude the long silences from watermarking, not to solve
the synchronization problem for watermark detection. Thus, the accuracy of locating
watermarking regions is not required to be as high as that in synchronization.
Similar to the method in [4], the proposed scheme selects the watermarking
regions according to the signal energy. The energy is calculated on a frame-by-frame
basis along the input signal, where each non-overlapping P frame g .n/ has a length
of N . Then, successive frames whose energy, i.e., N 2
nD1 g .n/, exceed a certain
threshold ET are concatenated to construct a high-energy segment. In our scheme,
a high-energy segment which has a duration of more than 2 s is considered to be a
long high-energy segment. Furthermore, adjacent long high-energy segments which
are located within 0.1 s are concatenated into one watermarking region.
The predefined energy threshold ET plays an important role in the selection of
watermarking regions. For different values of ET , different watermarking regions
are selected from the input signal, as shown in Fig. 4.1.
With a lower ET , more audio frames are included and hence longer segments
will be available for embedding the watermark. In this case, data payload is higher.
However, since the segments with relative low energy are susceptible to attacks, the
obtained watermarking regions might be unstable. With a higher ET , more stable
segments will be chosen for watermarking, but data payload is reduced accordingly.
Therefore, given the watermark to be embedded, ET is then determined to achieve
a better compromise between the robustness and data payload.
4.1 Preliminaries 97
Watermarking region 1 Watermarking region 2
0.7
−0.7
0 4 8 12 16 20 24
Time (s)
b
Region 1 Region 2 Region 3 Region 4

0.7
−0.7
0 4 8 12 16 20 24
Time (s)
Fig. 4.1 Selection of watermarking regions (a) ET D 0:01 (b) ET D 0:1
Frequency
fs Bsub(1) Bsub(2) Bsub(3) Bsub(4)
2
Block Block Block Block ... ...
0 Time
Fig. 4.2 Diagram of blocks in the watermarking domain
4.1.2 Structure of the Watermarking Domain
The watermarking domain, which is generated by taking the fast Fourier transform
(FFT) of adjacent audio frames with 50 % overlap, refers to the time–frequency
representation of the selected watermarking regions. Each frame has a length of N
points. As shown in Fig. 4.2, the watermarking domain is divided into Nblock blocks.
The reason for the half overlapping is to smooth the transition between frames, as
used by the previous techniques in Chap. 3.
Each block is used to embed one sub-watermark Bsub , which is a part of the
original watermark wo , i.e.,
˚
wo D Bsub.m/ m D 1; : : : ; Nblock : (4.1)
Subband Unit Frame Slot

(Frequency)
. ... .. . .......
...........
......
......
........
. ... .. . .......
. ... .. . .......
...........
......
...........
......
......
6
5 FFT coefficient
. .......
. .......
. .......
. .......
. .......
.. . . . . ..
Tile
4 (T12,4)
. . . ..
. . . ..
. . . ..
3
2 Subband
..
..
..
..
..
..
..
..
1
Frame index
0 1 2 3 4 5 6 7 8 9 10 11 12 (Time)
Fig. 4.3 Configuration of one block
Moreover, every sub-watermark Bsub contains Nbit watermark bits Bi 2 f1; 1g,
i.e.,
Bsub D fBi g i D 1; : : : ; Nbit : (4.2)
Figure 4.3 shows the details of each block as segmented into different levels of
granularities, such as unit, subband, slot, and tile.
Along the time axis, every block is divided into Nunit units, each of which
comprises Nc frames. Thus, one block has Nf D Nc Nunit frames. In our scheme,
Nc D 4. The concerns about the effect of Nc D 4 will be recognized in Sect. 4.3.1.
Along the frequency axis, the block is divided into Nsubband nonuniform perceptually
motivated subbands based on the Gammatone filterbank (GTF) (to be described in
Sect. 4.1.3).
The intersection of a subband and a unit is called a slot. Each watermark bit
Bi 2 f1; 1g is repeatedly embedded into a number of slots for robustness purposes.
In addition to the watermark bits, a synchronization bit Bs D 1 is also repeatedly
embedded for synchronization purposes. Specifically, NB slots are randomly chosen
from the total number of slots within every block for embedding each Bi . Then the
remaining slots are used for embedding Bs , whose total number is Ns . The value of
Ns is calculated by
Ns D Nunit Nsubband Nbit NB : (4.3)
In this way, every slot is used for embedding a bit, as exemplified in Fig. 4.4.
Without loss of generality, it is assumed here that each sub-watermark consists of
two watermark bits, i.e., Bsub D fB1 ; B2 g, which are separately embedded into five
slots, i.e., NB D 5. Thereby, Ns D 3625 D 8 slots are used for embedding Bs .
To enhance watermarking security, every bit embedded in the slot is modulated
by a pseudorandom number (PRN), Px 2 f1; 1g. As mentioned above, NB slots
4.1 Preliminaries 99
Subband
.. . .. ... . . . . . .. .
. . . . . . . . . . .. . . . .
. . . . . . . . . . .. . . . .
. . . . . . . . . . .. . . . .
.. . .. ... . . . . . .. .
.. . .. ... . . . . . .. .
. . . . . . . .. . . . . . . .
. . . . . .. . . . . . . .. .
. . . . . .. . . . . . . .. .
. . . . . . . . . . .. . . . .
. . . . . . . . . . .. . . . .
. . . . . .. . . . . . . .. .
6
4
3
2
1
0 1 2 3 4 5 6 7 8 9 10 11 12 Frame index
Fig. 4.4 Distribution of the watermark bits and synchronization bit. Note: Slots, , and
are used for embedding B1 , B2 , and Bs , respectively
are chosen for embedding each Bi . Accordingly, these NB slots correspond to

NB PRNs, denoted by Pbi D fPbi .k/g, k D 1; 2; : : : ; NB . Similarly, Ns slots
for embedding Bs correspond to a total of Ns PRNs, denoted by Ps D fPs .k/g,
k D 1; 2; : : : ; Ns .
The distribution of the bits in all blocks is determined by one secret key kb ,
which belongs to confidential information shared only between the embedder and
the authorized detectors. According to kb , different bits are spread on every block,
as exemplified in Fig. 4.4. To clearly describe the distribution of the bits, we build
up a bit matrix MB . Based on the structure of the block, MB is a Nsubband by Nunit
matrix. A PRN matrix MP of the PRNs is constructed correspondingly. The PRN
matrix MP is of the same size as MB , i.e., Nsubband Nunit . For example, MB and
MP for the block shown in Fig. 4.4 are expressed as
2 3 2 3
B1 Bs B2 Pb1 .3/ Ps .5/ Pb2 .5/
6B Bs 7 6 P .1/ Ps .8/ 7
6 2 B1 7 6 b2 Pb1 .4/ 7
6 7 6 7
6B Bs B2 7 6 P .2/ Ps .4/ Pb2 .4/ 7
MB D 6 1 7 and MP D 6 b1 7: (4.4)
6 Bs B2 Bs 7 6 Ps .2/ Pb2 .3/ Ps .7/ 7
6 7 6 7
4 Bs Bs B1 5 4 Ps .1/ Ps .3/ Pb1 .5/ 5
B1 B2 Bs Pb1 .1/ Pb2 .2/ Ps .6/
In this example, MP consists of Pb1 D fPb1 .1/ ; Pb1 .2/ ; : : : ; Pb1 .5/g, Pb2 D
fPb2 .1/ ; : : : ; Pb2 .5/g, and Ps D fPs .1/ ; Ps .2/ ; : : : ; Ps .8/g, which correspond to
B1 (in red), B2 (in blue), and Bs (in black), respectively. Without loss of generality,
our implemented watermark detection (to be described in Sect. 4.3) searches Bs and
each Bi which are embedded in the block, from the left bottom and then column
by column. Correspondingly, the indices of Ps and Pbi in Eq. (4.4) separately start
from the left bottom and column by column.
Note that all blocks have the same configuration of MB , which is solely
determined by kb . But the value of MB varies from block to block, since each block
has different watermark bits fBi g, i D 1; : : : ; Nbit to be embedded. As for MP ,
0.8
Amplitude frequency response
0.6
0.4
0.2
0
1000 2000 3000 4000 5000 6000 7000 8000
Frequency (Hz)
Fig. 4.5 Frequency response of a 16-channel GTF
it is unique and the same as all blocks to keep synchronization. The values of MP
components are determined by one secret key kp , which also belongs to confidential
information shared only between the embedder and the authorized detectors.
The smallest basic element of the slot is called tile as indicated in Fig. 4.3. A tile
consisting of several FFT coefficients is the basic module for amplitude modulation
in the watermark embedding. Due to Nc D 4, each slot contains four tiles. Recall
that every slot is used for embedding one bit, which is modulated by a PRN. More
specifically, the four tiles in a slot are used for embedding a bit and they share the
PRN of that slot in common. Accordingly, the tiles used for embedding each bit are
identified in the watermark detection and used to determine the bit value.
Each tile is denoted as Tt;b , where t and b are its frame and subband indices
respectively. For example, T12;4 shown in Fig. 4.3 stands for the tile located at the
4th subband of the 12th frame.
4.1.3 Gammatone Auditory Filterbank
As mentioned in Sect. 4.1.2, the Gammatone filterbank (GTF) is employed for

setting the subbands. The GTF is a bank of overlapping band-pass filters, which
mimics the frequency response of the human cochlea, wider bandwidths at higher
frequencies and narrower bandwidths at lower frequencies [124]. For the sake of
illustration, Fig. 4.5 shows an example of the frequency response of a 16-channel
GTF, covering 100–8,000 Hz frequency band.
4.2 Watermark Embedding 101
To get the center frequency and the bandwidth of each subband, the entire
carrying band (fL fH ) and the number of channels NGTF are required. Then,
overlapping spacing is calculated by

9:26 fH C 228:7
D log ; (4.5)
NGTF fL C 228:7
where fL and fH are lower and upper frequency limits in Hz.

So the nth .1 n NGTF / subband’s center frequency (fc ) and bandwidth (Bw )
in Hz are computed by
n
fc D 228:7 C .fH C 228:7/ exp (4.6)
9:26
Bw D 24:7 .1 C 4:37fc / : (4.7)
However, the channels from the GTF are usually overlapped, which results in
confusion of embedding bits. In order to get a set of non-overlapping subbands,
the lower limit of each channel (V l ) is always taken as the boundary, where V l D
fc Bw =2. This is because critical bandwidths are determined by the lower edges
of the band [11]. Moreover, narrow channels in the low frequencies might contain
only one FFT coefficient or even less under a sampling frequency of 44.1 kHz.
Since a single frequency coefficient is sensitive towards slight modification, several
channels in the low frequencies are combined in our scheme. In this way, the tiles
are forced to contain more than five FFT coefficients, which greatly helps improve
the robustness. For example, a set of 32 nonuniform subbands ( i.e., Nsubband D 32)
over the frequency spectrum is illustrated in Appendix G.
4.2 Watermark Embedding
Watermark embedding is to modulate the amplitude of the host audio signal in

the watermarking domain by a certain information used for copyrights protection
[125]. In this section, the embedding algorithm integrated with multiple scrambling
is described in detail.
4.2.1 Embedding Algorithm
For a given host signal to be watermarked, the first step of the embedding algorithm
is the selection of the watermarking regions. This is followed by the construction
of the watermarking domain. Generally, the watermark bits are embedded through
amplitude modulation of the tiles in the watermarking domain.
+ Watermarked
frame
+
Host frame Psychoacoustic Magnitude
Model 1 Watermark
signal
Phase
FFT Reconstruction IFFT
Frequency domain Time domain

Bit matrix Multiplication Sign
PRN matrix & Spreading
Fig. 4.6 Block diagram of watermarking one host frame
For each host frame, an imperceptible watermark signal is constructed in the

frequency domain using the magnitude, phase, and sign of the signal spectrum. Then
it is added to the host frame after being inversely transformed to the time domain.
Figure 4.6 shows the block diagram of watermarking one host frame.
The magnitude spectrum of the watermark signal is determined by Psychoacous-
tic Model 1. As discussed in Sect. 2.4, the input to the psychoacoustic model is one
host frame and the output is the minimum masking threshold (MMT). Therefore,
the noise introduced by the watermarking is kept inaudible to human ears.
It is worth mentioning that the MMT shown in Fig. 2.24 cannot be directly used
as the magnitude spectrum, because sound pressure level (SPL) normalization is
involved in the first step of implementing Psychoacoustic Model 1. Thus, we need
to multiply the MMT by a scale factor called watermark strength ˛w , in order to
obtain a proper magnitude spectrum. The effect of ˛w will be discussed in Sect. 5.1.
The phase of the watermark signal is the same as that of the host signal to avoid
phase distortion.
The sign of the spectrum of the watermark signal depends on the watermark
bits (i.e., bit matrix, MB ) and the corresponding PRNs (i.e., PRN matrix, MP ). As
described in Sect. 4.1.2, every slot is used for embedding a bit, which is modulated
by a PRN. So the sign of a slot is defined as the multiplication of its bit and PRN.
Then, the sign of a slot is spread over its four tiles. Specifically, if the sign of a slot
is positive, the sign of the first two tiles in that slot is positive and that of the last two
tiles is negative, and vice versa. Furthermore, the FFT coefficients in a tile share the
sign of that tile. In this way, all the FFT coefficients in the blocks obtain their own
signs.
After the calculation of the magnitude, phase, and sign, the frequency spectrum
of the watermark signal is constructed. Recall from Sect. 3.2.2.1 that a real-valued
signal has a conjugate symmetric spectrum. Thus, the construction process is
performed as follows:
%Fwm .1 W N / is the frequency spectrum of the watermark signal.

4.2 Watermark Embedding 103

N
%Fp 1 W is the positive-frequency part of Fwm :
2

N
%Fn 1 W is the negative-frequency part of Fwm :
2

N
Fp 1 W D sign .* magnitude .* exp (j*phase)I
2

N N
Fn 1 W D fliplr conj Fp 1 W I
2 2

N N
Fwm .1 W N / D Fp 1 W ; 0; Fn 1 W 1 I
2 2
where N is the frame length, function fliplr ./ is to flip a matrix left to right, and
function conj ./ is to compute the complex conjugate.
After taking the inverse fast Fourier transform (IFFT), the frequency spectrum of
the watermark signal is transformed to the time domain and subsequently added to
the host frame to produce the watermarked frame.
Finally, all the watermarked frames in each watermarking region are windowed
by a Hanning window for smooth concatenation. By combining original samples
with the watermarked regions in order, the overall watermarked audio signal is
formed.
Figure 4.7 shows an example of the watermarked signal, where the host signal
is the audio test file A2 (Bass.wav). It is observed that the watermark signal has a
similar shape to the host signal, which might help preserve the perceptual quality of
the watermarked signal.
4.2.2 Multiple Scrambling
To increase the level of security, multiple scrambling can be employed in the

embedding.
As described in Sect. 4.1.2, the amount and position of the slots used for
embedding each watermark bit are randomly set, where two secret keys, i.e., kb
and kp , are used. Furthermore, instead of using all subbands, we randomly select
NQ subband out of Nsubband subbands and randomize their orders of encoding, where
two secret keys are used. So the number of possible embedding ways is calculated
by the following permutation:
Nsubband Š
Nscrambling D P Nsubband ; NQ subband D : (4.8)
Nsubband NQ subband Š
a
0.7
−0.7
0 4 8 12 16 20 24
b
0.05
−0.05
0 4 8 12 16 20 24
c
0.7
−0.7
0 4 8 12 16 20 24
Time (s)
Fig. 4.7 Host signal and a watermarked signal by the proposed scheme. (a) Host signal.
(b) Watermark signal. (c) Watermarked signal
Another possible scrambling operation is to encrypt the coded-image watermark

(to be described in Sect. 4.4) into incomprehensible ciphers, where one secret key
is used. Even if the watermark is extracted, the attacker still cannot recognize the
image without the correct decipher.
Therefore, anyone without all the secret keys rarely has the ability to discern the
embedded watermark. Since the secret keys are shared only between the embedder
and authorized detectors, the aim of copyrights protection is really achieved. It is
worth mentioning that public-key algorithms in modern communication systems
[126] may be introduced to establish secure communication for sharing the secret
keys.
4.3 Watermark Detection
From the description of the embedding algorithm in Sect. 4.2.1, the watermark bits
in a bit matrix term are used to determine the sign of the spectrum of the watermark
signal. Therefore, in the detection, every watermark bit is determined by checking
4.3 Watermark Detection 105
whether it is an increase or decrease in the magnitudes of the corresponding tiles1

of the attacked signal.2 By correlating to the PRN matrix MP defined in Sect. 4.1.2,
we can detect the watermark without resorting to the host audio signal.
Since the watermark bits are embedded on a block-by-block basis, the beginning
frame of each block is required by the watermark detection to identify the tiles
used for embedding each bit. From the synchronization perspective, the beginning
frame in each block denotes the synchronization position of that block. However,
various attacks may modify the watermarked signals in different ways, and hence
the synchronization positions are destroyed. Therefore, synchronization methods
are required to find out the best synchronization position, which closely indicates
the beginning of each block.
The synchronization in our scheme is initially achieved by block synchronization.
As described in Sect. 4.1.2, a synchronization bit Bs D 1 is repeatedly embedded in
each block. Therefore, the synchronization position is considered to be the location
where the distribution of the tiles used for embedding Bs is matched.
However, as described in Sect. 4.1.2, the entire PRN matrix MP and the config-
uration of the bit matrix MB are the same for all blocks. Thus, in some cases when
the match with the tiles used for embedding Bs is obtained, the synchronization
position found by block synchronization is incorrect for detecting the watermark
bits. By introducing a threshold Tsync to check whether a synchronization position
can be accepted, adaptive synchronization [127] is developed to amend the block
synchronization.
As concluded from the study on different audio watermarking techniques, it
is more difficult to combat desynchronization attacks, especially pitch-invariant
time-scale modification (PITSM) and tempo-preserved pitch-scale modification
(TPPSM). Both can be implemented by audio editing tools without large per-
ceptual impairment.3 On most occasions, basic detection updated with adaptive
synchronization is robust to a limited extent against PITSM and TPPSM. To
efficiently handle excessive PITSM and TPPSM, frequency alignment [128] adjusts
the frequency spectra that have been scaled by the attacks, so that the detection can
retrieve the synchronization positions for recovering the embedded watermark.
Figure 4.8 shows the block diagram of watermark detection. The complete
algorithm consists of three parts: basic detection, adaptive synchronization, and
frequency alignment.
1
As mentioned in Sect. 4.1.2, the tile is the basic module for amplitude modulation in the
watermark embedding. Therefore, the detection is focused directly on the tiles, not the slots.
2
As defined in Sect. 1.2.1, the input to the watermark detector is generally called the attacked
signal, no matter whether it has been attacked or not. In the case that a watermarked signal has
been attacked, we specifically call it an attacked watermarked signal.
3
The random stretching attack used by [9, 10] which was implemented by omitting or inserting
a random number of samples (usually called “random samples cropping/inserting”) and the pitch
shifting attack by linear interpolation are much less complicated than PITSM and TPPSM.
Attacked signal
Basic detection
Selection of
wming regions
Watermarking
region
Frequency
Construction of spectrum Frequency
abs(.)
wming domain alignment
Magnitude
spectrum
Power
spectrum
Mean removal 20log10(.)
Descaled magnitude
spectrum
Whitened power
spectrum
Amplified
Inter- difference Analysis of Magnitude of
subtraction the tiles the tiles
Average
synchronization
strength (As)
Secret keys
(kb & kp)
Block Adaptive
synchronization synchronization
Synchronization position
(dsync)
Threshold
Bit (Tsync)
Bit strength Calculation of
determination bit strength
Watermark
bit
Secret keys
Watermark (kb & kp)
reconstruction
Final watermark
Fig. 4.8 Block diagram of watermark detection. Notes: (1) Basic detection works independently.
(2) Adaptive synchronization is an improvement technique for block synchronization. (3) Fre-
quency alignment indicated by dashed lines is an additional solution to excess PITSM and TPPSM
• Basic detection is the basic algorithm for watermark detection, which includes a
set of steps for detecting the watermark from the attacked signal. It begins with
the selection of watermarking4 regions and the construction of the watermarking
domain, similar to the embedding algorithm. Then, several operations are
performed to calculate the magnitude of the tiles, which are required by block
synchronization. By use of the secret keys, block synchronization aims to find
out the synchronization position of each block. Based on the synchronization
position found, the tiles used for embedding each watermark bit are identified
and further utilized to calculate the bit strength. According to the bit strength, the
value of that watermark bit is determined. Finally, the watermark bits detected
from all the blocks comprise the whole watermark.
• Adaptive synchronization is an improvement technique for block synchroniza-
tion. Except that a threshold Tsync is introduced as an extra input, adaptive
synchronization has the same inputs (i.e., magnitude of the tiles and secret keys)
and the same output (i.e., synchronization position dsync ) as block synchroniza-
tion. Unless requested by frequency alignment, adaptive synchronization does
not output the average synchronization strength (As ).
• Frequency alignment is an additional solution to excessive PITSM and TPPSM of
up to ˙10 %. Only when basic detection updated with adaptive synchronization
cannot detect the watermark from the attacked watermarked signal, frequency
alignment is employed to descale the frequency spectra. Average synchronization
strength from adaptive synchronization is involved in choosing the scale factor.
4.3.1 Basic Detection
As seen from Fig. 4.8, basic detection consists of the following eleven steps:
• Step 1: Selection of the watermarking regions
For a given signal, successive frames which exceed an energy threshold ET are
specified and concatenated into high-energy segments as watermark regions. Only
these segments are used for watermark detection.
• Step 2: Construction of the watermarking domain
Similar to the approach in the embedding, the watermarking domain for detection
is also generated by taking the FFT of adjacent frames with 50 % overlap. However,
Hanning windowing is applied on these frames before taking the FFT. This is
because all the watermarked frames are windowed and subsequently concatenated
into the watermarked signal, as described in Sect. 4.2.1.
The frequency spectrum of the tth windowed frame, Ft .n/, is calculated by
Ft .n/ D FFT fgt .n/ Hanning .n/g 1 n N; (4.9)
4
The word “watermarking” is abbreviated as “wming” in the first two boxes in Fig. 4.8.
where N is the frame length, gt .n/, 1 n N is the tth frame, and Hanning .n/,
1 n N is the N -point Hanning window.
In view of the conjugate symmetry of the frequency spectrum, we only process
the positive-frequency part with N2 coefficients, i.e., Ft 1 W N2 .
• Step 3: Calculation of the magnitude spectrum
ˇ ˇ
The magnitude spectrum of the tth windowed frame is shown to be ˇFt 1 W N
2
ˇ,
where jj is the absolute value.
• Step 4: Calculation of the power spectrum
The power spectrum of the tth windowed frame, Pt 1 W N
2
, is calculated by
ˇ ˇ
N ˇ N ˇˇ
Pt 1 W ˇ
D 20 log10 ˇFt 1 W : (4.10)
2 2 ˇ
• Step 5: Calculation of the whitened power spectrum

The whitened power spectrum,POt 1 W N
2
, is calculated by removing the mean from
the power spectrum:

O N N
Pt 1 W D Pt 1 W Pt ; (4.11)
2 2
1 X
N=2
where Pt D ŒPt .n/ is the mean of Pt 1 W N
2
.
N=2
nD1
• Step 6: Inter-subtraction
To amplify the effect of the watermark signal and reduce the effect of the host
signal, the difference PQt 1 W N2 between each frame and the one just after the next
is calculated:

N N N
PQt 1 W D POt 1 W POtC2 1 W : (4.12)
2 2 2
The reason is that on the condition that adjacent frames in the blocks are half
overlapped, these two frames (t and t C 2) are considered to be non-overlapped
with each other.
• Step 7: Analysis of the tiles
To accumulate the effect of the watermark signal over the FFT coefficients in a
tile, the magnitude of each tile is calculated. Specifically, the magnitude of the tile
located at the bth subband of the tth frame, Qt;b , is calculated by
Vb h
X
PQt .n/
nDVbl
Qt;b D ; (4.13)
Vbh Vbl C 1
where Vbl and Vbh refer to the lower and upper bounds of the bth subband,
respectively.
• Step 8: Block synchronization
The purpose of block synchronization is to find out the beginning frame of each
block.
Since every frame in its block is possibly the beginning frame, it is necessary
to calculate synchronization strength Sd d D 1; : : : ; Nf frame by frame, where
Nf D Nc Nunit . On the assumption that the dth frame is the beginning frame, Sd
for the dth frame is calculated by the following normalized correlation [13]:
X
Ns

Qt.d;k/; b.k/ Ps .k/
kD1
Sd D v v ; (4.14)
u Ns u Ns
uX uX
t Qt.d;k/; b.k/
2
t ŒPs .k/2
kD1 kD1
˚
where Qt.d;k/; b.k/ represent Ns tiles that are used for embedding Bs under the
assumed condition and fPs .k/g are their corresponding PRNs. The subscripts
t .d; k/ and b .k/ to locate the tiles are computed by
t .d; k/ D d C ŒRs .k; 1/ 1 Nc (4.15)

b.k/ D Rs .k; 2/; (4.16)
where Nc is the number of frames per unit and Nc D 4 is considered in our scheme.
Rs is called the index matrix for Bs , which solely depends on the secret key kp
mentioned in Sect. 4.1.2. To indicate the location of these tiles, the first and second
columns of Rs represent the distribution of the tiles’ columns and rows, respectively.
For example, given MB in Eq. (4.4) which corresponds to Fig. 4.4, its Rs is shown
as follows. As mentioned already in Eq. (4.4), the search for Bs starts from the left
bottom and column by column without loss of generality.
(4.17)
Then, the beginning frame of this block (i.e., the synchronization position dsync )
is the frame that provides the maximum Sd :
dsync D arg max.Sd /: (4.18)

1d Nf
Also, the maximum Sd of that dsync is denoted as Sdsync .

• Step 9: Calculation of bit strength
Based on the synchronization position found by Eq. (4.18), the bit strength of each
watermark bit Bj , Gj , is calculated by
NB h
X i
Qt .dsync ;k /; b.k/ Pbj .k/
kD1
Gj D v v ; (4.19)
u NB h i2 u
uX uXNB

t Qt .dsync ;k /; b.k/ t Pbj .k/
2
kD1 kD1
where the subscripts are computed by

t dsync ; k D dsync C Rbj .k; 1/ 1 Nc (4.20)
b .k/ D RBj .k; 2/ : (4.21)
n o ˚
Similarly, Qt .dsync ;k /; b.k/ and Pbj .k/ refer to NB tiles that are used for
embedding Bj and their corresponding PRNs, respectively. As for Rbj , it is the
index matrix for Bj . Also, based on MB of Fig. 4.4, Rb1 and Rb2 are built as follows:
2 3 2 3
11 15
61 47 62 17
6 7 6 7
6 7 6 7
Rb1 D 6 1 6 7 and Rb2 D 6 2 3 7 :
6 7 6 7
42 55 43 45
32 36
Cropped frames
Subband
Block Block Block
Frame index
1 4 8 12 1 5 10 12
Fig. 4.9 Illustration of random samples cropping
• Step 10: Determination of the watermark bit

The value of watermark bit, Bj , is determined according to its bit strength Gj .
If Gj 0; then Bj D 1:
(4.22)
If Gj < 0; then Bj D 0:
• Step 11: Reconstruction of the watermark

The watermark bits extracted from all the blocks are combined and the final
watermark we is obtained.
4.3.2 Adaptive Synchronization
Adaptive synchronization is an improvement technique for block synchronization.

As described in Sect. 4.3.1, the watermark bits are detected based on the synchro-
nization position dsync found by block synchronization. Recall from Sect. 4.1.2
that the distribution of the synchronization bit Bs (indexed by Rs ) and the values of
the corresponding PRNs (i.e., Ps ) are the same for all blocks. Thus, in some cases,
although the dsync from Eq. (4.18) provides the best match with the tiles used for
embedding Bs , it is a false position for detecting the watermark bits.
Take random samples cropping as an example of the attacks. Suppose that the
first three frames in the second block have been cropped, as illustrated in Fig. 4.9.
Our task is to detect the watermark bits embedded in the second block. To this
end, block synchronization is performed for this block, which means that the 12
frames indicated with red numbers are separately assumed to be the beginning
frame. Due to the fact that these three blocks have the same Rs and Ps , the dsync
found from Eq. (4.18) would be the current tenth frame that is originally the first
frame of the third block. Then, based on Eqs. (4.20) and (4.21), the tiles for bit
Block synchronization
Magnitude of
the tiles (Qt,b) Calculate Sd dsync Yes
Determine dsync ~
(d =1,…, Nf) dsync Tsync dsync
Secret keys based on Eq.(3.18)
by Eq.(3.14)~(3.16)
(kb & kp)
No
Replace Eq.(3.15)
with Eq.(3.23)
Fig. 4.10 Flowchart of adaptive synchronization
detection in Eq. (4.19) will be taken from the third block, instead of the second
block. Consequently, the watermark bits embedded in the second block cannot be
correctly detected.
To find out the actual synchronization position of the second block (i.e., the
current first frame), we need to compensate for the cropped frames. According to
the structure of the block, the number of the cropped frames can be calculated by
dm D Nf dsync C1 D Nc Nunit dsync C1 D 4310C1 D 3. Therefore, when
identifying the tiles for embedding Bs , we take into consideration an offset of dm .
Specifically, t .d; k/ in Eq. (4.15) should be modified into the following equation:
t .d; k/ D d C ŒRs .k; 1/ 1 Nc dm ; (4.23)
where dm D Nf dsync C 1. Subsequently, new tiles are identified for block

synchronization, in order to find out the actual synchronization position dQsync of
this block.
One point to note is that t .d; k/ calculated by Eq. (4.23) might be a negative
value. For example, suppose that MB in Eq. (4.4) is embedded in the second block
in Fig. 4.9, then Rs is given in Eq. (4.17) and dm D 3. According to Eq. (4.23),
t .1; 1/ and t .1; 2/ are calculated by t .1; 1/ D 1 C ŒRs .1; 1/ 1 4 3 D 2
and t .1; 2/ D 1 C ŒRs .2; 1/ 1 4 3 D 2. In this case, these tiles cannot be
identified and their values are simply replaced by zeros.
Furthermore, t dsync ; k in Eq. (4.20) for calculating Gj is also modified into

t dQsync ; k D dQsync C Rbj .k; 1/ 1 Nc dm : (4.24)
The procedure described above is called adaptive synchronization. The key of

adaptive synchronization is to choose a threshold Tsync for determining whether
a dsync can be accepted. A dsync is considered to be incorrect if it is larger than
Tsync . Then, dsync should be recalculated and eventually reach a value lower than
Tsync , which will be taken as the dQsync . Figure 4.10 shows the flowchart of adaptive
synchronization.
Note that if more than half of the frames in a block are missing or destroyed, we
do not expect to detect the watermark bits embedded in that block. Therefore, 1
N N
dm 2f is considered. In view of dm D Nf dsync C1, then 2f C1 dsync Nf .
This means that under different attacks, Tsync is varied between 1 C Nf =2 and Nf .
Thus, we need to perform adaptive synchronization for each possible Tsync and sub-
sequently calculate the average synchronization strength Async that is defined as
1
NX
block h i
Async D SdQsync .k/ ; (4.25)
Nblock
kD1
where SdQsync .k/ is the kth block’s SdQsync . Then, the Tsync that provides the maximum
Async is ˘regarded as the desired one. Experimentally, an optimal value of Tsync is
0:8Nf , where bc is the smallest integer value.
4.3.3 Frequency Alignment Towards Excessive PITSM

and TPPSM
Among various attacks, time-scale modification (TSM) and pitch-scale modification

(PSM) are more likely to raise difficulties in the process of watermark detection. In
most cases, adaptive synchronization can only cope with PITSM and TPPSM at
a limited extent (within ˙4 %). Although the requirement of the SDMI standard
described in Appendix A has been satisfied, a higher level up to ˙10 % is desired
by STEP 2000 described in Appendix B.
In order to combat an excessive distortion of PITSM and TPPSM, we suggest
performing frequency alignment to adjust the frequency spectra that have been
scaled. Hence, synchronization positions can be retrieved for recovering the embed-
ded watermark from a severely attacked watermarked audio file.
4.3.3.1 Frequency Alignment Against TSM and PSM
Time- and pitch-scale modification of audio signals refer to the operations of

independently controlling and modifying the time evolution and the pitch contour,
respectively. In particular, PITSM is to slow down or speed up a given signal without
altering its pitch. TPPSM, meanwhile, is to shift the pitch upward or downward
without affecting the duration [129].
As their definitions imply, there is a duality between TSM and PSM [129]. A
pitch-scaled signal can be obtained by a TSM followed by a sampling rate alteration.
On the other hand, a time-scaled signal can be achieved by changing the sampling
rate of a signal after PSM, as shown in Fig. 4.11. This shows that a solution to
withstanding PSM is also applicable to TSM, since a time-scaled signal can be
Positive TSM downsampling Positive PSM

(slower tempo) upsampling (higher pitch)
Negative TSM upsampling Negative PSM

(faster tempo) downsampling (lower pitch)
Fig. 4.11 Duality between TSM and PSM
converted to a pitch-scaled signal after an alteration of the sampling rate. Therefore,

PSM is used for illustrating how to perform frequency alignment.
The operation of PSM can be formulated as
fQ D ˛ f; (4.26)
where ˛ .> 0/ is the scale factor, f is the frequency of the original signal, and fQ is
the ˛-scaled frequency [112]. If ˛ > 1, it is a positive PSM that gets a higher pitch.
If ˛ < 1, it is a negative PSM that gets a lower pitch. Accordingly, such a frequency
fluctuation introduces desynchronization in watermark detection.
To retrieve the synchronization positions, frequency alignment attempts to reverse
the process of PSM. Thus, the modified frequency spectrum is descaled by
falignment D ˇ fQ; (4.27)
where ˇ D 1=˛. It appears as an operation of compression for positive PSM and

expansion for negative PSM. In this manner, the original f could be recovered
approximately.
4.3.3.2 Implementation of Frequency Alignment
The strategy for conducting frequency alignment in watermark detection is

described as follows. After applying basic detection updated with adaptive
synchronization on an attacked signal, we calculate the bit error rate (BER) between
the original watermark .wo / and the extracted watermark .we /. If the BER is less
than the threshold TBER , the attacked signal under inspection is claimed to be a
watermarked copy and the detection process is terminated [2]. If the BER is larger
than TBER , it is considered that the suspected signal might have been attacked by
excessive TSM and/or PSM. Thus, frequency alignment is employed to descale
the frequency spectra before the calculation of magnitude spectra, in an attempt to
recover the synchronization positions.
It is worth mentioning that TBER is not necessary to be an accurate value
and merely provides a rough idea of whether the detection proceeds or not.
Moreover, one valuable benefit offered by coded-image watermark (to be described
in Sect. 4.4) is a semantic meaning of copyright information. The detection is
a
iF 1 2 3 4 5 6 7 8 ...
Ft Ft (1) Ft (2) Ft (3) Ft (4) Ft (5) Ft (6) Ft (7) Ft (8) ...
iX 1 2 3 3 4 4 4 ... ...
Average
Average

Dt Ft (1) Ft (2) Ft 3, 4 Ft 5, 6, 7 ... ... ... ...
b
iF 1 2 3 4 5 6 7 8 ...
Ft Ft (1) Ft (2) Ft (3) Ft (4) Ft (5) Ft (6) Ft (7) Ft (8) ...
iX 1 2 3 5 8 ... ... ... ...
Dt Ft (1) Ft (2) Ft (3) Ft (4) Ft (5) ... ...
Linear interpolation
Fig. 4.12 Illustration of frequency alignment (a) Positive PSM (b) Negative PSM
terminated provided that the copyright information can be recognized. Even without
knowing TBER , we can always perform frequency alignment on any attacked signal,
regardless of the attacks which are unknown beforehand. But frequency alignment
probably would fail to extract the watermarks modified by other attacks rather
than TSM/PSM attacks. Eventually, if the watermarks extracted by all attempts are
unrecognized, the suspected signal is claimed to be unwatermarked. Therefore, the
proposed audio watermarking scheme is undoubtedly blind.
Figure 4.12 shows the process of frequency alignment, where ˛ is a scale factor,
Ft is the frequency spectrum of the tth windowed frame in Eq. (4.9), and Dt is the
resulting vector after performing frequency alignment. Note that we consider both
positive- and negative-frequency parts of Ft during the calculation; hence, the length
of Ft is N . However, only the first N2 points of Dt , i.e., Dt 1 W N2 , will be taken as
the descaled frequency spectra for further processing.
In general, the elements of Ft are indexed by a vector iF , i.e., iF D 1 W N .
Another index vector iX is calculated by rounding the result of iF =˛ to the nearest
integer, i.e., iX D round .iF =˛/. The values of iX in turn determines the resulting
vector, Dt .
Specifically, for a positive PSM, ˛ > 1 may result in repetitive values of iX ,
i.e., the nth, .n C 1/th, . . . , and .n C x/th elements have the same value. If the nth
element of iX has a unique value m, then transfer Ft .n/ to Dt .m/. For instance,
in our example in Fig. 4.12a, the first element of iX has a unique value of 1, this
indicates transferring Ft .1/ to Dt .1/. Otherwise, calculate the average of the nth to
.n C x/th elements (which correspond to repetitive values of iX ) and transfer this
value to Dt .m/. For instance, in our example, the third and fourth elements of iX
both have values of 3, which indicates transferring the average of Ft .3/ and Ft .4/
to Dt .3/.
As for negative PSM (˛ < 1), the values of iX are discontinuous. If the nth
element of iX has a value of n, then transfer Ft .n/ to Dt .n/. For instance, in our
example in Fig. 4.12b, the first element of iX equals 1, this indicates transferring
Ft .1/ to Dt .1/. If the nth element of iX has a value m, where m ¤ n, then transfer
Ft .n/ to Dt .m/, e.g., in our example, the fourth element of iX equals 5, this indicates
transferring Ft .4/ to Dt .5/. As the values of iX is discontinuous, only part of the
vector Dt is filled by this mechanism. The rest of the vector is calculated by linear
interpolation between successive known values.
As shown above, the outcomes of the frequency alignment for a positive and
negative PSM are different—one being compression and the other being expansion.
Using this knowledge, we can use one positive and one negative trial values of the
scale factor respectively to descale the frequency spectra. Two trial values are gener-
ally within ˙10 %, e.g., C6 % and 6 %. After the frequency spectra are descaled,
adaptive synchronization is performed to output the average synchronization
strength Async . Let us denote the value of Async obtained from the positive trial value
by AC C
sync and that from the negative trial value by Async . If Async > Async , it can be
deduced that the watermarked signal has been attacked by a positive PSM, and vice
versa. Further, the scale factor can be delicately adjusted for a higher detection rate.
To convert PITSM to TPPSM, the length of the host signal .No / is required to
be the information shared between the embedder and the detector. By comparing
with the length of the attacked signal .Na /, it is ascertained whether the attack
is a positive or negative PITSM. Accordingly, the attacked signal is resampled to
the corresponding TPPSM. Then, frequency alignment is performed to improve the
accuracy of watermark detection. Although a slight deviation of Na might occur
(which happens when samples cropping or inserting attack the watermarked signal
along with PITSM), experimental results in the next chapter show that such an
amount of difference is negligible to the operation of resampling.
4.3.3.3 Error Analysis Associated with TBER
As mentioned in Sect. 4.3.3.2, the BER threshold TBER is used to determine whether
one detection is successful or not and further to declare whether a watermark exists
or not. In practice, we are given a signal under inspection and then perform water-
mark detection to extract a supposed watermark. Once one extracted watermark has
a BER less or equal than TBER , the suspected signal is claimed to be watermarked.
Otherwise, the suspected signal is claimed to be unwatermarked if all attempts to

detect the watermark always get BERs more than TBER .
Consequently, there are two types of errors in determining the existence of a
watermark, i.e., a false-positive error and a false-negative error. A false-positive
error (or false alarm) occurs when the detector indicates the presence of a watermark
in an unwatermarked signal, and the false- positive probability is the likelihood of
such an occurrence. A false- negative error (or miss detection) occurs when the
detector indicates the absence of a watermark in a watermarked signal, and the false-
negative probability is the likelihood of such an occurrence [13]. It is rather difficult
to establish exact models for both false-positive and false-negative probabilities.
Here, simple binomial models [2, 5, 130] are used in our analysis: the extracted Nw
bits are assumed to be Nw independent, identically distributed Bernoulli variables
with the same “success” probability p, where Nw is the watermark length. Note that
a “success” means the extracted bit matching the original watermark bit.
According to the definition, a false-positive error of one detection occurs if this
detection extracts .Nw Ne / or more bits successfully, where Ne D bNw TBER c
is the number of wrong bits. Since the “success” probability p of each bit extracted
from an unwatermarked signal is supposed to be 12 , the false-positive probability of
one detection Ppd is calculated as follows [2, 5, 130]:
X k
1 .Nw k/
Nw
1
Ppd D C .Nw ; k/ 1
2 2
kD.Nw Ne /
(4.28)
1 X
Nw
D C .Nw ; k/;
2Nw
kD.Nw Ne /
where C .Nw ; n/ denotes the number of combinations, i.e., C .Nw ; n/ D nŠ.NNwwn/Š

Š
.
Furthermore, since the suspected signal is claimed to be watermarked as soon as
one detection is successful, the false-positive probability of watermark existence is
determined by [2, 130]
X
Ndet
n .Ndet n/
Ppw D C .Ndet ; n/ Ppd 1 Ppd ; (4.29)
nD1
where Ndet is the total number of detections that were performed.

The analysis of false-negative probability is more complicated than false-positive
probability, because false-positive probability depends on the watermark detection
algorithm only; however, false-negative probability depends on both watermark
embedding and detection algorithms. Moreover, since the watermarked signal might
be distorted by various attacks, an accurate false-negative probability should be
calculated for a specific attacked signal [13]. In our case, a general model is
employed to roughly estimate the false-negative probability.
The false-negative probability is computed in a different way from the

false-positive probability, i.e., calculating the correct detection probability which
is merely the complement of the false-negative probability. Considering that
the “success” probability p of each bit extracted from an attacked signal is
approximated to be .1 TBER /, the correct detection probability of one detection
Pcd is calculated as follows [130]:
X
Nw
Pcd D C .Nw ; k/ .1 TBER /k .TBER /.Nw k/ : (4.30)
kD.Nw Ne /
Recall that the suspected signal is claimed to be watermarked if at least one detection
is successful. Therefore, the false-negative probability of watermark existence is
determined by [130]
X
Ndet
Pnw D 1 C .Ndet ; n/ .Pcd /n .1 Pcd /.Ndet n/ : (4.31)
nD1
The severity of false-positive and false-negative probabilities is application depen-

dent. Mostly, more concern is focused on minimizing the occurrence of false-
positive errors. For the application of copyrights protection, a very low false-positive
probability (less than 105 ) is of a higher priority [5, 130], and meanwhile a
moderate false-negative probability is desired under various attacks within tolerable
levels.
4.4 Coded-Image Watermark
From the analysis of the embedding and detection algorithms, the watermark
embedded for copyrights protection is essentially represented by a series of
watermark bits.
To better function for copyrights protection, the proposed scheme adopts coded-
image such as with bit “1” and “0” (mapped to “1”)5 as a visual
watermark, instead of a meaningless pseudorandom or chaotic sequence. A coded-
image can be identified visually, as a kind of ownership stamp. Moreover, post-
processing on the extracted watermark can also be done to enhance the binary image
and consequently the detection accuracy will increase. Image denoising and pattern
recognition are examples of post-processing techniques for automatic character
recognition. Thus, on top of BER, coded-image provides a semantic meaning for
reliable verification [131]. In addition, as mentioned in Sect. 4.2.2, the watermarking
scheme benefits from the encryption of the coded-image to obtain extra security.
5
By definition, a coded-image belongs to a binary image, which has only two values for each pixel.
4.4 Coded-Image Watermark 119
Fig. 4.13 Coded-image Noise B

denoising by morphological Noise C
operations
Noise A
In our scheme, a coded-image watermark consists of letters, each of which is

represented by a matrix of 7 5 bits. It is worth mentioning that the coded-image
watermark is embedded letter by letter, not line by line. In this way, even when a
part of the watermarked signal is severely attacked, the letters in the other parts still
can be clearly recognized.
Discussion of post-processing techniques on binary images is beyond the scope
of this thesis. For completeness, two possible methods are briefly described below,
i.e., morphology and neural network (NN) [132].
Morphology is a technique of image processing based on shapes. The fundamen-
tal morphological operations are erosion and dilation, which can be used in a variety
of ways to give other transformations including opening, closing, skeletonization,
and so on. A simple application of the opening operation is to remove small objects
with fewer pixels than a threshold from the image. For example, noise B and C in
Fig. 4.13 can be easily eliminated by this means. However, it is impossible to get rid
of noise A which is connected with the informative letters. In this case, skeleton-
based character recognition might be a good solution to retrieve the distorted
coded-image. But character skeletonization is not quite applicable to the proposed
scheme, since the line width of the coded-image does not contain a large number of
pixels.
Neural networks have been widely used for character recognition. In this book, a
backpropagation neural network from the Neural Network Toolbox for MATLAB6
is employed to recognize all 26 capital letters of the alphabet. Each letter is
represented as a 7 by 5 matrix, such as the letters shown in Fig. 4.14a. As illustrated
in Fig. 4.14b, c, noisy letters on the extracted coded-image watermark could be fully
recovered by the neural network.
6
Appcr1: Character Recognition at http://www.mathworks.com/access/helpdesk/help/toolbox/
nnet/.
b c
Fig. 4.14 Character recognition by the neural network. (a) Letters “C” “O” “P” “Y” “R” “I” “G”
“H” “T”. (b) Noisy coded-image watermark. (c) Recovered coded-image watermark
4.5 Summary
As discussed in Chap. 3, spread spectrum technique contributes to the robustness

and security of audio watermarking schemes. However, traditional SS watermarking
is likely to suffer from perceivable watermark embedding and desynchronization
attacks. In this chapter, an imperceptible, robust, and secure audio watermarking
scheme using Psychoacoustic Model 1, multiple scrambling, adaptive synchroniza-
tion, frequency alignment, and the coded-image watermark is proposed. Essentially,
a watermark modulated by PRNs is embedded in the time–frequency domain of the
host signal. Accordingly, watermark detection is based on normalized correlation
between the watermarked signal and corresponding PRNs.
To overcome the problem of perceivable watermark embedding, only the regions
with high energy are selected to prevent noise caused by watermarking the silence.
Moreover, a half-overlapped Hanning window and the GTF involved in the con-
struction of the watermarking domain are also beneficial to transparent perception.
Most significantly, the MMT from Psychoacoustic Model 1 is utilized to shape the
amplitude of the watermark signal, so that the watermarks are embedded without
noticeably degrading the perceptual quality of the audio signals.
Meanwhile, several measures have been taken to cope with various attacks,
especially desynchronization attacks. Commonly, basic detection with block syn-
chronization is applied to seek the positions of synchronization bits as well as
watermark bits. However, the synchronization positions found are considered incor-
rect when exceeding a certain threshold. In such cases, adaptive synchronization is
employed to search for the correct synchronization positions. Furthermore, to resist
severe PITSM and TPPSM of up to ˙10 %, frequency alignment is developed to
descale the distorted frequency spectra of the attacked watermarked signal. Thus,
the synchronization positions can be retrieved for successful watermark detection.
On top of imperceptibility and robustness, the proposed scheme is also strictly
self-secured by using multiple scrambling. It is extremely difficult for any attacker
lacking all the secret keys to ascertain or destroy the embedded watermark. In
addition to the security benefited from the encryption, the usage of the coded-image
4.5 Summary 121
watermark as a visual identification makes it possible to improve the accuracy of

watermark detection further by employing image processing techniques and pattern
matching analysis.
This chapter has presented a theoretical analysis of the embedding and detection
algorithms. In Chap. 5, the performance of the proposed audio watermarking
scheme will be thoroughly evaluated through perceptual quality assessment, robust-
ness test and security analysis, etc.
Chapter 5
Performance Evaluation of Audio
Watermarking
In Chap. 4, the embedding and detection algorithms of the proposed audio water-
marking scheme were analyzed theoretically. The aim of this chapter is to examine
system performance in terms of imperceptibility, robustness, security, data payload,
and computational complexity, as required in Sect. 1.3.1.
First, the process of determining the parameters used for watermarking is
described. Then performance measurement begins with perceptual quality assess-
ment, which consists of the subjective listening test and the objective evaluation test.
This is subsequently followed by a complete robustness test including both basic and
advanced robustness tests. After performing a security analysis, we carry out the
estimations of data payload and computational complexity. Finally, a performance
comparison is made between the proposed scheme and other existing schemes.
Some observations are discussed according to the experimental results.
5.1 Experimental Setup
To investigate the performance of the proposed audio watermarking scheme, a

series of experiments were carried out on different audio signals. It is worth
mentioning that all the audio signals in the test set are taken as host audio signals
in order to inspect the applicability of the proposed scheme. But for the sake
of illustration, we choose one typical audio signal from each category only and
present their simulation results separately in the following sections. Nevertheless,
similar results can be found for other signals. The selected audio files are (i) vocal,
Bass:wav .A2 /; (ii) percussive instruments, Glockenspiel2:wav .A7 /; (iii) tonal
instruments, Harpsichord:wav .A8 /; and (iv) music, Pop:wav .A15 /. For simplicity,
Glockenspiel2.wav and Harpsichord.wav are shortened as Gspi.wav and Harp.wav
respectively.
Recall that audio watermarking always involves the trade-off relationships among
imperceptibility, robustness, security, data payload, computational complexity,
and so on. Thus, experiment parameters should be properly set to optimize the
124 5 Performance Evaluation of Audio Watermarking
system performance for the intended application. As mentioned in Sect. 1.2.3.1,

imperceptibility, robustness, and security are the key criteria in designing any
audio watermarking scheme for copyrights protection. Accordingly, the parameters
involved in the watermarking are determined based on their requirements.
• Determination of the variables N , Nunit , Nsubband , Nbit , and NB
The variables associated with the structure of the watermarking domain defined
in Sect. 4.1.2 are first identified, including the frame length .N /, the number of
units per block .Nunit /, the number of nonuniform subbands .Nsubband /, the number
of watermark bits embedded in one block .Nbit /, and the number of slots for
embedding each watermark bit .NB /.
Note that the value of frame length is fixed at N D 512 due to the use of
Psychoacoustic Model 1, and hence N2 D 512 2
D 256 FFT coefficients per frame
are available for watermarking. Moreover, as mentioned in Sect. 4.1.3, each tile is
required to contain more than five FFT coefficients for the purpose of robustness.
Following the calculations for the channels of the Gammatone filterbank, a proper
value of Nsubband is obtained as Nsubband D 32. On considering the operation of
multiple scrambling, the number of selected subbands NQ subband is set to be 28. Then,
the values of Nbit and NB are subsequently decided. So with Nunit , Nsubband , Nbit ,
and NB , the number of slots for embedding the synchronization bit .Ns / can be
calculated by Eq. (4.3) as shown below:
Ns D Nunit NQ subband Nbit NB (5.1)
These variables combine to affect the system performance in some way. Accord-
ing to the embedding algorithm previously described, larger Ns and NB could
contribute to a stronger robustness against desynchronization attacks to some extent.
Correspondingly, a larger value of Nunit and a smaller value of Nbit are desired.
However, a larger Nunit would lead to an increase in computational complexity.
As indicated in Eq. (4.18), the number of times that each block searches for its
synchronization position is equal to Nf D Nc Nunit , where Nc D 4. Also, the
values of all tiles in the blocks should be provided simultaneously for watermark
detection, thus more computer memory is required to store the data. Additionally,
as a result of a smaller Nbit , data payload would be inevitably reduced.
In view of these constraints, the above variables are specified as follows to aim for
a good compromise between various requirements. That is, Nunit D 10, Nbit D 4,
NB D 30, and the resulted Ns D 160, which will be employed as constants for all
experiments.
• Determination of watermark strength ˛w
The amplitude of the watermark signal in the embedding has an important influence
on system performance. As discussed in Sect. 4.2.1, the magnitude spectrum of
the watermark signal is controlled by the watermark strength ˛w . To get good
performance in the experiments, watermark strength might be selected to be
uniformly distributed between 10 and 200, i.e., ˛w D 10; 20; 30; : : : ; 200.
5.1 Experimental Setup 125
Given a host audio signal, watermarking with a smaller ˛w would result in better
imperceptibility, but weaker robustness; on the other hand, watermarking with a
larger ˛w would result in imperceptibility degradation, but stronger robustness [19].
Considering that imperceptibility is a prerequisite for practical application of audio
watermarking, our prime concern in determining ˛w is dedicated to maintaining the
perceptual quality of the watermarked audio signals. Within the scope of satisfactory
perceptual quality, a value of ˛w that provides adequate robustness is adopted. In our
experiments, the software “Perceptual Evaluation of Audio Quality” (PEAQ) [48]
is used to interpret the perceptual quality, and an objective difference grade (ODG)
within Œ2:0; 0 is deemed to be acceptable. Also, the property of robustness is
denoted by the bit error rate (BER) of the extracted watermark under 36 dB noise
addition attack. 1 Empirically, a reasonable BER is expected to be less than 10 %.
For example, Fig. 5.1 shows the determination of watermark strength ˛w for
Bass:wav and P op:wav. As indicated in Fig. 5.1a, watermarking with ˛w 50
is considered unperceived, since the ODGs fit within the allowable range Œ2:0; 0.
Also, under the condition of ˛w 50, the BERs are less than 10 % and thereby the
requirement of robustness is met. Consequently, ˛w D 50 is the only appropriate
value for watermarking Bass:wav. By comparison, embedding a robust watermark
into P op:wav while retaining the imperceptibility is more feasible. With the same
method, it is found that a proper watermark strength for watermarking P op:wav
ranges between 60 and 140, i.e., 60 ˛w 140. In this case, we use the average
value ˛w D 100 for the experiments with P op:wav below.
• Determination of the embedded watermark
Generally, the less watermark bits embedded, the better imperceptibility but the
worse robustness. To evaluate the proposed scheme fairly, our experiments always
embed the watermark bits into host signals at full capacity. For example, the number
of watermark bits .Nw / embedded into Bass.wav, Gspi.wav, Harp.wav, and Pop.wav
is 350, 210, 140, and 280, respectively. More analysis of data payload will be
presented later in Sect. 5.5.1.
As shown in Sect. 4.4, a coded-image watermark offers greater advantage over
a pseudorandom sequence (PRS). Therefore, the coded-image watermark is always
adopted in the experiments. Recall that each letter on the coded-image watermark
is represented by a matrix of 7 5 bits, i.e., Lw D 35 bits for one letter. Thus,
Nw watermark bits can be coded into NL D Nw =Lw letters. For example, the
coded-image watermark embedded into Bass.wav is , which consists of
NL D 350=35 D 10 letters; the coded-image watermark embedded into Gspi.wav
is , which consists of NL D 210=35 D 6 letters; the coded-image watermark
1
Additive noise attack is a commonly used attack in robustness test of audio watermarking
techniques. As clearly indicated in Appendix A and B, SDMI standard and STEP 2000 employ
36 dB and 40 dB additive noise attack respectively. Therefore, a rigorous additive noise attack with
a lower SNR value, i.e., 36 dB additive noise attack, is chosen for our basic robustness test listed
in Appendix E.
a
ODG
0 40%
BER
−0.5 35%
−1 30%
Imperceptibility (ODG)
ODG = −1.986
Robustness (BER)
−1.5 BER = 9.14% 25%
−2 20%
−2.5 15%
−3 10%
−3.5 5%
−4 0%
0 50 100 150 200

Watermark strength αw
b
ODG
0 40%
BER
−0.5 35%
−1 30%
Imperceptibility (ODG)
Robustness (BER)
−1.5 25%
−2 20%
−2.5 15%
ODG = −1.076
BER = 3.93%
−3 10%
−3.5 5%
−4 0%
0 50 100 150 200

Watermark strength αw
Fig. 5.1 Determination of watermark strength. ˛w (a) ˛w D 50 for Bass:wav. (b) ˛w D 100 for
P op:wav
5.2 Perceptual Quality Assessment 127
embedded into Harp.wav is , which consists of NL D 140=35 D 4 letters; and

the coded-image watermark embedded into Pop.wav is , which consists of
NL D 280=35 D 8 letters.
Based on the above parameters, the audio signals in the test set are watermarked.
Then, the watermarked Bass.wav, Gspi.wav, Harp.wav, and Pop.wav signals are
used for illustration in the following sections, except for the test under collusion in
Sect. 5.3.3.2 and the test under multiple watermarking in Sect. 5.3.3.3. In the tests
under collusion and multiple watermarking, a number of coded-image watermarks
are embedded into each host signal as required.
5.2 Perceptual Quality Assessment
The goal of perceptual quality assessment is to fairly judge the perceptual quality
of the watermarked audio signals relative to host audio signals. To this end, both
subjective and objective approaches to perceptual quality assessment are employed
in this book, as discussed in Sect. 1.3.2.1.
5.2.1 Subjective Listening Test
Subjective listening tests are carried out in two ways: the MUSHRA test and the five-
scale subjective difference grade (SDG) rating. As described in Sect. 3.1.2, listening
tests performed in an isolated chamber were undertaken by ten trained participants
and all the stimuli are presented through a high-fidelity headphone.
The MUSHRA test stands for MUlti Stimuli with Hidden Reference and Anchors
test, which is defined by ITU-R recommendation BS.1534 [44]. In the MUSHRA
test, the participant is exposed to three types of audio clips as test, reference (i.e.,
the original unprocessed audio), and anchor audio signals. The recommendation
specifies that one anchor must be a 3.5 kHz low-pass filtered version of the reference
audio signal [44]. Also, a hidden reference is usually adopted as another anchor.
Then, the participant is asked to grade the perceptual quality of the audio signals
under test and the anchors relative to the reference audio signal.
We developed a MATLAB GUI for the MUSHRA test to help our analysis, as
shown in Fig. 5.2. In the context of audio watermarking, the watermarked signal
is the signal under test, while the host signal is the reference signal. As required,
the host signal is always presented in the experiments. For the anchors, we use three
versions of the host signal, i.e., a hidden version, a 3.5 kHz low-pass filtered version,
Fig. 5.2 Screenshot of the MATLAB GUI for the MUSHRA test. The buttons on the GUI have
the following functions: “Load,” load the host audio signal to be evaluated. “Start,” start playing
the sound from the beginning. “Pause/Stop,” pause or stop the sound that is currently playing.
“Resume,” resume the sound from the pause position. “Save,” save the host signal name and the
participant name as well as the registered scores into a .txt file. “Reset,” reset the interface for the
next trial
and a 96 kbps MP3 compressed version.2 During each experiment, the participant is
therefore asked to grade four versions of a given host signal, i.e., the watermarked
signal (WM), the hidden reference (HOST), the low-pass filtered version (LPF), and
the compressed version (MP3).
Given a host signal, one participant can launch the MUSHRA test by clicking
the “Load” button. Subsequently, four versions of the host signal will be randomly
assigned to Source AD. Then, the participant grades each version by moving the
slider to the location corresponding to the perception. Accordingly, a score between
Œ0; 100 appears in the text box below. With the buttons “Start,” “Pause/Stop,”
and “Resume,” the participant can switch instantly between different sound files.
Finally, the buttons “Save” and “Reset” save the test results and get ready for next
experiment.
2
The 3.5 kHz low-pass filtered version refers to a version of host audio filtered by a 3.5 kHz low-
pass filter, and the 96 kbps MP3 compressed version refers to a version of host audio after MP3
compression at 96 kbps.
As mentioned in Sect. 3.1.1, the test set includes 17 pieces of audio signals,
denoted by Ai , i D 1; 2; : : : ; 17. Then, the four versions of Ai are denoted by
Aij , where j D 1; 2; 3, and 4 stands for the version of WM, HOST, LPF, and MP3,
respectively. Since there are 10 subjects participating in the tests, the score of Aij
provided by the k-th subject is denoted by GM .i; j; k/, where k D 1; 2; : : : ; 10.
After all the scores are collected, statistical analysis [44] is performed to assess
the perceptual quality of each Aij separately. First, the mean of the scores of Aij is
calculated by
1 X
K
ij D GM .i; j; k/ (5.2)
K
kD1
where K D 10 in our experiments.

Then, a 95 % confidence interval .˛ D 1 0:95 D 0:05/ about the mean value
ij is given by

ij ıij ; ij C ıij (5.3)
where
ij
ıij D t0:05 p (5.4)
K
Here, t0:05 is the t test for a significance level of 95 % and is the standard deviation
defined as
v
u
u 1 XK

t 2
ij D GM .i; j; k/ ij (5.5)
.K 1/
kD1
On the assumption that the mean scores follow a normal distribution, the value of
t0:05 is equal to
˛
t0:05 D ˆ1 1 D ˆ1 .0:975/ D 1:96 (5.6)
2
where ˆ1 ./ is the inverse normal cumulative distribution function.

After substituting t0:05 in Eq. (5.6) into Eq. (5.4), the 95 % confidence interval in
Eq. (5.3) becomes

ij 0:62 ij ; ij C 0:62 ij (5.7)
Figure 5.3 shows the results of statistical analysis on Bass.wav, Gspi.wav,

Harp.wav, and Pop.wav. Different versions of each host signal are spread along the
Bass.wav
100
Gspi.wav
90 Harp.wav
Pop.wav
80
70
MUSHRA score
60
50
40
30
20
10
0
WM HOST LPF MP3
Fig. 5.3 Results of the MUSHRA-based subjective listening test
X-axis, while the perceptual quality scores are lying along the Y-axis. The average
scores over ten listeners as well as the related 95 % confidence interval are displayed.
From Fig. 5.3, the hidden references are rated the highest with small 95 %
confidence intervals. On the other hand, the scores of 3.5 kHz low-pass filtered
signals are quite low. This is because a low-pass filter attenuates high frequencies,
which makes audio samples sound dull. Moreover, the scores of watermarked
signals are comparable to those of MP3-compressed signals at 96 kbps, which
means they are of similar perceptual quality. For different host audio signals,
the watermarked Harp.wav signal has the best performance and its average score
is around 93. The second is the watermarked Pop.wav signal, followed by the
watermarked Gspi.wav signal. Although the watermarked Bass.wav signal obtains
the lowest average score, the score is still more than 80. Therefore, we conclude that
perceptual quality of all the watermarked signals is well preserved. Also, perceptual
quality depends on the music type. It is worth mentioning that three observations are
common issues among all techniques, due to the complexity of audio signals [133].
In addition to MUSHRA test, a rating based on the five-scale SDG evaluates
the perceptual quality of the watermarked signals in a straightforward manner. The
subjects are asked to rate a watermarked signal relative to its host signal according
to the descriptions in Table 1.2. Similarly, the SDG of host signal Ai from the k-th
subject is denoted as GSDG .i; k/, where i D 1; 2; : : : ; 17 and k D 1; 2; : : : ; 10.
Then, the average SDG of host signal Ai is calculated as
1 X
K
i D GSDG .i; k/ (5.8)
K
kD1
where K D 10.
Table 5.1 Results of the SDG-based subjective listening test

Bass.wav Gspi.wav Harp.wav Pop.wav
Average SDG 1:52 1:27 0:45 0:83
Table 5.2 Results of the objective evaluation test

ODG 1:986 1:758 0:509 1:076
SNR/dB 33:39 32:01 24:44 30:43
For example, the average SDGs for the watermarked Bass.wav, Gspi.wav,
Harp.wav, and Pop.wav signals are shown in Table 5.1. From the table, it is seen
that the average SDGs for these samples are in the range of 1:6 and 0. In fact,
except that a few feel a slight difference between Bass.wav and its watermarked
signal, most listeners find it hard to distinguish the host and the watermarked audio
signals during the experiments.
5.2.2 Objective Evaluation Test
As discussed in Sect. 3.1.2, objective evaluation tests on the watermarked audio

signals include two metrics, namely the ODG by using software PEAQ and the
signal-to-noise ratio (SNR) defined in eq. (1.3).
The values of the ODG and the SNR for the watermarked Bass.wav, Gspi.wav,
Harp.wav, and Pop.wav signals are shown in Table 5.2. As can be seen, it is
acceptable that the ODGs are within (2:0; 0) below the threshold of becoming
slightly annoying. By averaging these four values, the average ODG of the proposed
scheme is calculated to be 1:33. Moreover, the ODG values accord with the results
of the subjective listening test. That is, the watermarked Harp.wav signal is the
most imperceptible, followed in descending order by the watermarked Pop.wav,
Gspi.wav, and Bass.wav signals.
In addition, the SNR values are higher than the 20 dB requirement from the
International Federation of the Phonographic Industry (IFPI) [2, 14]. By averaging
these four values, the average SNR value of the proposed scheme is calculated to be
30.1 dB.
From Tables 5.1 and 5.2, one point to note is that the SNR values are not
in agreement with the actual perceptual quality in terms of the SDG. More
investigation on other objective quality measures will be made in the next chapter.
From the above results of perceptual quality assessment, it is verified that our
watermarked signals are mostly perceptually undistinguished from the host audio
signals.
5.3 Robustness Test
The goal of the robustness test is to investigate the capability of the watermarked
audio signals to resist various attacks. To fully evaluate the robustness of the
proposed audio watermarking scheme, we carry out both basic and advanced
robustness tests as depicted in Sect. 3.1.3. All ODGs in this section are provided
by using PEAQ.
5.3.1 Error Probability
Recall from Sect. 4.3.3.3 that the threshold TBER has an influence on both false-
positive (Ppw ) and false-negative (Pnw ) probabilities of declaring the existence of a
watermark. Suppose that the value of TBER is equal to 20 % [2, 5], Ppw and Pnw are
calculated using Eqs. (4.28)–(4.31).
Table 5.3 shows the results on the error probabilities of the watermarked
Bass.wav, Gspi.wav, Harp.wav, and Pop.wav signals, where Nw is the watermark
length and Ne D bNw TBER c is the number of wrong bits. As mentioned in
Sect. 5.1, 350, 210, 140, and 280 watermark bits are separately embedded into
Bass.wav, Gspi.wav, Harp.wav, and Pop.wav, so the resulting Ne is equal to 70, 42,
28, and 56, respectively. In addition, Ndet is the number of detections performed,
where two values (i.e., Ndet D 5 and 10) are considered.
For different host signals in Table 5.3, the false-positive probabilities Ppw
increase exponentially with the watermark length Nw , but vary slightly with
the number of detections performed Ndet . On the other hand, the false-negative
probabilities Pnw increase with Ndet , but vary slightly with Nw .
Generally, the severity of false-positive and false-negative probabilities is applica-
tion dependent. In our scheme, given TBER D 20 %, the false-positive probabilities
have already satisfied the requirement, being much less than 105 . When Ndet D 5
and 10, the false- negative probabilities are around 102 and 104 , respectively.
These values are sufficient for the application of copyrights protection [130].
Therefore, the threshold is set to be TBER D 20 % in our experiments. This means
that the detections with the BERs of greater than 20 % are considered failed.
Table 5.3 Results of error probabilities under TBER D 20 %

Nw 350 210 140 280
Ne 70 42 28 56
Ndet D 5 Ppw 1:78 1030 1:20 1018 1:06 1012 1:44 1024
Pnw 2:25 102 2:03 102 1:84 102 2:16 102
Ndet D 10 Ppw 3:57 1030 2:40 1018 2:12 1012 2:88 1024
Pnw 5:05 104 4:14 104 3:38 104 4:66 104
5.3 Robustness Test 133
5.3.2 Basic Robustness Test
As described in Sect. 3.1.3.1, a variety of common signal operations and desyn-

chronization attacks are included in the basic robustness test, for example, noise
addition, resampling, requantization, amplitude scaling, low-pass filtering, echo
addition, reverberation, MP3 compression, DA/AD conversion, random samples
cropping, jittering, zeros inserting, pitch-invariant time-scale modification (PITSM),
and tempo-preserved pitch-scale modification (TPPSM). Moreover, a number of
combined attacks are constructed to test the robustness, such as random samples
cropping, jittering, or zeros inserting followed by low-pass filtering or MP3
compression. Also, PITSM or TPPSM followed by low-pass filtering, MP3 com-
pression, random samples cropping, jittering, or zeros inserting are employed to
attack the watermarked signals, in order to verify the validity of frequency alignment
under challenging conditions.
As described in Sect. 4.3, the whole procedure of watermark detection begins
with basic detection. Then, basic detection updated with adaptive synchronization
(called improved detection, for simplicity) is always performed to improve the
detection rate. In cases of PITSM and TPPSM, frequency alignment is further used
to descale the frequency spectra with an attempt to retrieve the distorted watermark.
For simplicity, improved detection integrated with frequency alignment is called
advanced detection. Note that it is not always necessary to proceed with frequency
alignment to combat PITSM and TPPSM attacks. Empirically, if the improved
detection has extracted a watermark with a BER of less than 10 %, such a result
is considered to be already satisfactory for detection.
Tables 5.4, 5.5, 5.6, and 5.7 show the results of the basic robustness test on
the watermarked Bass.wav, Gspi.wav, Harp.wav, and Pop.wav signals, respectively.
Recall that the coded-image watermarks embedded are for Bass.wav,
for Gspi.wav, for Harp.wav, and for Pop.wav.
To illustrate the effectiveness of adaptive synchronization and frequency align-
ment, the BERs of the extracted watermarks at each stage of the detection are
separately presented. Specifically, the “Basic detection” column, the “Adaptive
synchronization” column, and the “Frequency alignment” column show the BERs
of the extracted watermarks obtained by basic detection, improved detection, and
advanced detection, respectively. Then, the extracted watermark obtained by the last
detection is taken as the final extracted watermark, we , illustrated in the last column.
As discussed in Sect. 5.3.1, TBER D 20 %. Accordingly, detections with BERs
of greater than 20 % are considered failed and indicated by the symbol “” in
the tables. Moreover, the symbol “” indicates the situation whereby advanced
detection with frequency alignment is unexecuted. This is because the BER of the
extracted watermark obtained by improved detection is already less than 10 %.
The following observations are obtained from the analysis of Tables 5.4, 5.5, 5.6,
and 5.7.
• On the whole, the proposed audio watermarking scheme demonstrates strong
robustness against various attacks, although different watermarked audio signals
Table 5.4 Results of the basic robustness test on the watermarked Bass.wav signal
Basic Adaptive Frequency Final
detection synchronization alignment watermark
(BER: %) we
No attack 0 0
SNR (dB) 40 dB 6.86 5.71
36 dB 11.71 9.14
30 dB 18.57
Resampling (22.05 kHz) 0 0
Requantization (8 bit) 17.71 16.00
Amplitude C20 % 0 0
20 % 0 0
Lp filtering 8 kHz 0 0
6 kHz 1.43 1.43
5 kHz 9.71 9.71
DA/AD (line-in jack) 0
Echo (0.3, 200 ms) 0.57 0.57
Reverb (1 s) 0 0
Compression II 96 kbps 0 0
64 kbps 1.43 1.14
48 kbps 3.71 2.86
Cropping (8 25 ms) 0
Jittering (0.1 ms/20 ms) 0
Inserting (8 25 ms) 0 0
PITSM C4% 0.57
C10 % 2.86
4 % 1.71
10 % 6.00
TPPSM C4 % 14.86 8.00
C10 % 4.86
4 % 12.57 6.00
10 % 4.29
Notes: 1. Symbol “”: one detection with a BER of greater than 20 %
2. Symbol “”: one unexecuted advanced detection
differ in the performance. In terms of the BER, almost all the BERs of we are
less than 10 %. Moreover, the coded-image watermarks embedded in the four
host signals can always be extracted and clearly identified. Therefore, compared
to merely meaningless bits, extra assistance in confirmation can be obtained
from the coded-image watermarks. Although some pixels in the coded-image
are mistaken, we are still able to recognize the copyrights information.
• For different watermarked signals, the watermarked Harp.wav signal in Table 5.6
shows the strongest robustness. None of the BERs are greater than 10 %, in fact
most of them are equal to 0 %. This is because a higher watermark strength ˛w
is used in watermarking Harp.wav, as can be seen from the lower SNR value of
the watermarked Harp.wav signal in Table 5.2. Due to the characteristics of the
Table 5.5 Results of the basic robustness test on the watermarked Gspi.wav signal
(BER: %) we
No attack 0 0
SNR (dB) 40 dB 12.38 7.62
36 dB 12.38 9.52
30 dB 16.67 16.19
Resampling (22.05 kHz) 1.90 1.90
Requantization (8 bit) 19.05 19.05
Amplitude C20 % 0 0
20 % 0 0
Lp filtering 8 kHz 0.95 0.95
6 kHz 6.67 4.29
5 kHz 11.43
DA/AD (line-in jack) 0
Echo (0.3, 200 ms) 0 0
Reverb (1 s) 0 0
Compression II 96 kbps 2.86 2.86
64 kbps 3.81 3.81
48 kbps 14.76 9.05
Jittering (0.1 ms/20 ms) 3.33
PITSM C4 % 0.48 0.48
C10; % 10.00
4 % 4.29
10 % 10.48 4.29
TPPSM C4 % 10.95 3.81
C10 % 9.52
4 % 12.86 4.29
10 % 11.90
harpsichord, Harp.wav can be watermarked with a higher ˛w on the premise of a

satisfactory perceptual quality.
This is followed by the watermarked Bass.wav signal in Table 5.4 and the
watermarked Pop.wav signal in Table 5.7. For the watermarked Bass.wav signal,
all the BERs of we are less than 10 %, except for the ones under 30 dB noise
addition and requantization attacks. Similarly, for the watermarked Pop.wav
signal, all the BERs of we are also less than 10 %, except for the ones under
30 dB noise addition, requantization, and C10 % TPPSM attacks. On average,
the watermarked Gspi.wav signal in Table 5.5 shows the weakest resistance to
the attacks. Four final watermarks, which are attacked by 30 dB noise addition,
Table 5.6 Results of the basic robustness test on the watermarked Harp.wav signal
(BER: %) we
No attack 0 0
SNR (dB) 40 dB 0 0
36 dB 0 0
30 dB 0 0
Resampling (22.05 kHz) 0
Requantization (8 bit) 0 0
Amplitude C20 % 0 0
20 % 0 0
6 kHz 10.00 0
5 kHz 15.71 10.00
DA/AD (line-in jack) 0 0
Echo (0.3, 200 ms) 0 0
Reverb (1 s) 0 0
Compression II 96 kbps 17.14 0
64 kbps 0
48 kbps 4.29
PITSM C4% 2.86 2.86
C10 % 2.14
4 % 0
10 % 0.71
TPPSM C4 % 6.43
C10 % 1.43
4 % 12.86 4.29
10 % 4.29
requantization, 5 kHz low-pass filtering, and 10 % TPPSM, have a BER of

greater than 10 %.
• For different attacks, all the watermarked signals exhibit rather high robustness
against most attacks. The improved detection can almost perfectly extract
the watermarks from the watermarked signals attacked by amplitude scaling,
resampling, DA/AD conversion, echo addition, reverberation, MP3 compression,
random samples cropping, jittering, and zeros inserting. It is worth mentioning
that our watermarked signals are quite robust against DA/AD conversion, where
the BERs of we are all equal to 0 %. However, as shown in Chap. 3, most audio
watermarking techniques cannot resist DA/AD conversion.
Table 5.7 Results of the basic robustness test on the watermarked Pop.wav signal
(BER: %) we
No attack 0 0
SNR (dB) 40 dB 0.71 0.71
36 dB 3.93 3.93
30 dB 12.14 10.71
Resampling (22.05 kHz) 0 0
Requantization (8 bit) 19.29
Amplitude C20 % 0 0
20 % 0 0
6 kHz 11.79 2.50
5 kHz 16.43 9.64
DA/AD (line-in jack) 13.93 0
Echo (0.3, 200 ms) 0 0
Reverb (1 s) 13.93 0
Compression II 96 kbps 0 0
64 kbps 0 0
48 kbps 2.50 2.50
PITSM C4 % 8.57 2.14
C10 % 9.64
4 % 1.07
10 % 15.36 4.29
TPPSM C4% 10.00 1.43
C10 % 12.14
4 % 14.29 0.36
10 % 8.93
In cases of PITSM and TPPSM attacks, the improved detection can mostly
combat PITSM and TPPSM within ˙4 %, but fail at larger distortions of
˙10 %. Under such circumstances, we resort to the advanced detection with
frequency alignment to extract the severely distorted watermarks. From Table 5.4
to Table 5.7, it can be seen that the BERs of we attacked by ˙10 % PITSM, and
TPPSM attacks are reduced greatly after frequency alignment, most of which are
not greater than 10 %.
Among various attacks, it is observed that requantization is the most difficult
attack. Except for the watermarked Harp.wav signal in Table 5.6, the other three
watermarked signals are rather vulnerable to the requantization and their BERs
are not less than 16 %. Moreover, noise addition poses difficulties for watermark
detection as the power of the added noise increases. Note that the decrease of
the specified SNR value, i.e., 40 dB ! 36 dB ! 30 dB, indicates the increase
of the noise power. To enhance the robustness against these two attacks, higher
watermark strengths ˛w are required to amplify the magnitude of the watermark
signals in the embedding.
In addition to common signal operations and the desynchronization attacks listed
above, some combined attacks are also applied on the watermarked signals. The
aim is to further evaluate the robustness of the proposed scheme. A combined
attack is a combination of two attacks, i.e., AT1 ./ followed by AT2 ./. The
procedure of applying a combined attack on the watermarked signal is as follows:
the watermarked signal is first attacked by AT1 ./ and the resulting signal is then
attacked by AT2 ./.
Two types of combined attacks are taken into consideration.
(1) Type I combined attack: AT1 ./ is random samples cropping, jittering, or zeros
inserting, while AT2 ./ is MP3 compression at 96 kbps or low-pass filtering at
8 kHz.
(2) Type II combined attack: AT1 ./ is C5 % PITSM, 5 % PITSM, C5 % TPPSM,
or 5 % TPPSM, while AT2 ./ is MP3 compression at 96 kbps, low-pass
filtering at 8 kHz, random samples cropping, jittering, or zeros inserting.
Without loss of generality, the watermarked Bass.wav signal is used as an
example. The coded-image watermark embedded into Bass.wav is
Table 5.8 shows the results of combined attacks on the watermarked Bass.wav
signal, including the final extracted watermarks we as well as their BERs. Note
that all the we attacked by Type I combined attacks are extracted by the improved
detection. Type II combined attacks are very destructive and the improved detections
fail to extract the watermarks. In this case, the advanced detection is employed to
recover the we attacked by Type II combined attacks.
In Table 5.8, each combined attack is the combination of the attacks on the
corresponding row and column. For instance, in (1) Type I combined attacks, the
shaded BER (0.86 %) and we are the results under the combined attack where AT1 ./
is random samples cropping and AT2 ./ is MP3 compression at 96 kbps. Also, in
(2) Type II combined attacks, the shaded BER (0.57 %) and we are the results under
the combined attack where AT1 ./ is 5 % PITSM and AT2 ./ is zeros inserting.
From Table 5.8, it can be seen that the proposed scheme is quite resistant to
these combined attacks. All the BERs of we are less than 10 % and the coded-
image watermarks can be clearly identified. It is observed that the combined attacks
involving jittering are generally more challenging than the others.
With regard to Type II combined attacks, one point to note is that the length
of the PITSM- or TPPSM-attacked signal has been altered by cropping, jittering,
and inserting. Even in these cases, the distorted watermarks can be recovered by
advanced detection. Therefore, it is proved that a slight change in the length of
the attacked signal has no influence on the efficiency of frequency alignment, as
discussed in Sect. 4.3.3.2.
Table 5.8 Results of combined attacks on the watermarked Bass.wav signal

(1) Type I combined attacks
Cropping, jittering, or inserting followed by compression or filtering
Compression II (96 kbps) Lp filtering (8 kHz)
BER: % we BER: % we
Cropping (8 25 ms) 0.86 1.43
Jittering (0.1 ms/20 ms) 8.86 9.14
Inserting (8 25 ms) 1.14 1.71
(2) Type II combined attacks
(a) PITSM followed by compression, filtering, cropping, jittering, or inserting
PITSM .C5 %/ PITSM .5 %/
BER: % we BER: % we
Compression II (96 kbps) 5.71 4.57
Lp filtering (8 kHz) 4.29 8.00
Cropping (8 25 ms) 1.43 1.71
Jittering (0.1 ms/20 ms) 9.43 6.00
Inserting (8 25 ms) 1.71 0.57
(b) TPPSM followed by compression, filtering, cropping, jittering, or inserting
TPPSM .C5 %/ TPPSM .5 %/
BER: % we BER: % we
Compression II (96 kbps) 5.14 4.00
Lp filtering (8 kHz) 7.71 7.71
Cropping (8 25 ms) 1.71 2.29
Jittering (0.1 ms/20 ms) 7.71 4.86
Inserting (8 25 ms) 2.00 1.14
5.3.3 Advanced Robustness Test
The advanced robustness test is designed especially for evaluating the proposed
audio watermarking scheme. As described in Sect. 3.1.3.2, the advanced robustness
test is comprised of three parts, namely a test with StirMark for Audio, a test
under collusion, and a test under multiple watermarking. Note that in the advanced
robustness test, all the watermarks are extracted by improved detection, i.e., the
basic detection updated with adaptive synchronization.
5.3.3.1 Test with StirMark for Audio
StirMark for Audio [134] is a publicly available benchmark for robustness evalu-
ation of audio watermarking schemes. In the experiments, we utilize StirMark for
Audio v0.2 with default parameters and a suite of 50 StirMark-attacked signals are
generated accordingly. Note that the attacked signals from StirMark for Audio are
stereo signals. Similar to the method in Sect. 3.1.1, the left channel is taken as the
attacked watermarked signal in our scheme.
Based on the description of the attacks in Appendix C and the analysis of
the attacked watermarked signals, the following attacks are excluded from the
evaluation: Addfftnoise, Extrastereo_30, Extrastereo_50, Extrastereo_70, Nothing,
Resampling, and Voiceremove. Since the audio test files are monaural, Extrastereo
attack has no effect. Also, as its name implies, the Nothing attack does nothing
with the watermarked signals. So in these cases, the watermarked signals remain
unchanged and the watermarks can always be extracted perfectly. Moreover, most
samples of the attacked signals under Addfftnoise, Resampling, and Voiceremove are
zeros, and hence it is unnecessary to proceed with the detection. Other than these,
the remaining 43 attacks are included in our experiments. It is worth mentioning that
the Original attack resembles the original (unattacked) watermarked signal, which
is actually the same as the Nothing attack. Therefore, the resulting signals from the
Original attack can be referenced as the original watermarked signals.
Detection results of the watermarked Bass.wav, Gspi.wav, Harp.wav, and
Pop.wav signals under StirMark for Audio are shown in Table 5.9. For simplicity,
the coded images are not illustrated in the table, and we only present the BERs of
the extracted watermarks. Apart from the BERs, the ODGs of the attacked signals
relative to their host signals are also calculated to get an insight into the amount of
distortion caused by the attacks. If the ODG of the attacked signal is comparable
to that of the original watermarked signal, it means that the watermarked signal is
less affected by this attack. Otherwise, the watermarked signal has already been
severely destroyed by the attack.
Table 5.9 shows that the proposed scheme has high resistance to most attacks in
StirMark for Audio, including common signal operations and some serious desyn-
chronization attacks, such as Copysamples, Zerolength, and Zeroremove. Although
more than half or all four watermark detections fail under the shaded attacks (i.e.,
Addnoise_500, Addnoise_700, Addnoise_900, Cutsamples, and Zerocross), these
attacks actually have strong negative impact on the fidelity of the watermarked
signals. As can be seen, in cases of failed detections, the ODGs are lower than
3:30, while 80 % of the ODGs are even lower than 3:80. Therefore, the attacked
signals are very different from the host signals, beyond the premise of the robustness
test.
For different watermarked signals, the watermarked Harp.wav signal possesses
the strongest robustness. Except for one failed detection and the detection under the
Echo attack, the rest of the attacks cannot destroy the embedded watermarks and the
resulting BERs are not more than 5 %. Next comes the watermarked Pop.wav signal,
which succeeds in 34 detections with BERs of less than 8 % and three detections
with BERs of around 16 %. For the watermarked Bass.wav signal, the BERs of all
35 surviving watermarks are less than 7 %. Similar to the conclusion in the previous
section, the watermarked Gspi.wav signal suffers more from the attacks. But even
so, the watermarked Gspi.wav signal fails in six detections only.
Finally, it should be pointed out that successful detections under Addbrumm_1100
Addbrumm_10100 and Addsinus attacks are on the condition that the initial
Table 5.9 Results of StirMark for Audio attacks

StirMark attacks BER: % ODG BER: % ODG BER: % ODG BER: % ODG
Original 0 1.986 0 1.737 0 0.509 0 1.076
Addbrumm_100 0 2.189 0 1.891 0 0.88 0 1.084
Addbrumm_1100 0 3.412 0 3.245 0 3.109 0 1.783
Addbrumm_2100 0 3.635 0 3.472 0 3.405 0 2.610
Addbrumm_3100 0 3.753 0 3.580 0 3.495 0 3.140
Addbrumm_4100 0 3.804 0 3.651 0 3.565 0 3.414
Addbrumm_5100 0 3.830 0 3.699 0 3.61 0 3.560
Addbrumm_6100 0 3.846 0 3.732 0 3.637 0 3.644
Addbrumm_7100 0 3.856 0 3.761 0 3.658 0 3.696
Addbrumm_8100 0 3.863 0 3.782 0 3.683 0 3.730
Addbrumm_9100 0 3.868 0 3.798 0 3.705 0 3.755
Addbrumm_10100 0 3.872 0 3.814 0 3.712 0 3.772
Addnoise_100 6.86 3.763 11.9 3.870 0 3.468 3.57 3.565
Addnoise_300 3.880 11.43 3.907 0 3.804 16.07 3.847
Addnoise_500 3.892 3.911 0 3.844 3.877
Addnoise_700 3.898 3.911 0 3.858 3.885
Addnoise_900 3.846 3.912 2.86 3.863 3.888
Addsinus 0 3.895 0 3.819 0 3.814 0 3.848
Amplify 0 2.921 0 2.613 0 1.772 0 2.623
Compressor 0 3.858 0 2.058 0 0.51 0 1.076
Copysample 3.883 5.71 3.913 1.43 3.642 3.624
Cutsamples 3.908 3.913 3.908 3.868
Dynnoise 3.798 3.908 0 3.14 16.07 3.855
Echo 3.898 12.86 3.891 17.14 3.867 15.36 3.872
Exchange 1.14 2.989 5.71 3.638 3.57 2.391 7.86 2.271
FFT_hlpass 0 3.835 6.19 3.873 0 3.373 0 3.206
FFT_invert 0 2.128 0 1.829 0 0.502 0 1.085
FFT_real_reverse 0 2.135 0 1.934 0 0.561 0 1.083
FFT_stat1 4.29 3.894 7.14 3.838 2.86 3.284 2.50 3.829
FFT_test 4.29 3.894 7.14 3.838 2.86 3.291 2.50 3.829
Flippsample 0.57 3.680 5.24 3.913 0 2.417 0.36 3.360
Invert 0 1.986 0 1.737 0 0.509 0 1.076
Lsbzero 0 2.011 0 1.775 0 0.545 0 1.097
Normalize 0 3.317 0 2.813 0 3.371 0 3.357
RC_highpass 0 2.435 0 2.092 0 1.114 0 1.727
RC_lowpass 0 2.205 0 1.759 0 0.868 0 1.456
Smooth 1.71 3.618 19.05 3.899 0 2.965 1.43 2.458
Smooth2 6.29 3.537 13.81 3.897 0 2.884 0 2.400
Stat1 0 1.897 0 1.776 0 0.58 0 0.654
Stat2 0 2.191 0 1.801 0 0.826 0 1.378
Zerocross 3.638 3.896 5.00 3.171 3.336
Zerolength 0.57 3.903 0.95 3.894 0 2.341 0 3.867
Zeroremove 0 3.911 0 3.909 0 3.888 0 3.659
Note: Symbol “”: one detection with a BER of greater than 20 %
watermarking regions are known to the detector. These attacks add high-amplitude
buzz or sinus tone throughout the watermarked signal3 , which has an influence on
the threshold ET for the selection of watermarking regions. As a result, the detector
cannot properly locate the regions for watermark detection. In this case, ET must
be set at a higher value to select more stable regions. However, as discussed in
Sect. 4.1.1, data payload is reduced accordingly.
5.3.3.2 Test Under Collusion
Collusion is one challenging statistical attack on audio watermarking schemes.

Given n watermarked signals sw.1/ ; sw.2/ ; : : : ; sw.n/ that are generated by separately
embedding n watermarks w.1/ .2/
o ; wo ; : : : ; wo
.n/
into host signal so , the collusion
attack is to create n average watermarked signals sw.i/ as follows:
8
<sw.j / D Embedding s ; w.j / ; 1 j n
o o
:s .i/ D 1 s .1/ C s .2/ C : : : C s .i/ ; 1 i n
(5.9)
w i w w w
.i;j /
In the detection, i watermarks we are detected from each average watermarked
signal sw.i/ individually:

w.i;j
e
/
D Detection sw.i/ ; 1 i n and 1 j i (5.10)
Such an averaging operation weakens the original watermarks and hence makes
them hard to detect. Note that the averaging collusion attack is the most common
collusion attack and nonlinear collusion attacks [135] are not taken into considera-
tion in the book.
During our experiments, four different watermarks (w.1/ .2/ .3/
o , wo , wo ;and wo )
.4/
are separately embedded into host signal so and yield four watermarked signals
(sw.1/ , sw.2/ , sw.3/ , and sw.4/ ). Then, four average watermarked signals are generated,

i.e., sw.1/ D sw.1/ , sw.2/ D 12 sw.1/ C sw.2/ , sw.3/ D 13 sw.1/ C sw.2/ C sw.3/ , and sw.4/ D
.1/
1
4
sw C sw.2/ C sw.3/ C sw.4/ .
After that, four sets of watermark detections are separately performed as fol-
lows:
(1) Detect w.1/ .1/ .1;1/
o from sw to obtain the extracted watermark we .
.1/ .2/ .2/
(2) Detect wo and wo separately from sw to obtain the extracted watermarks
we.2;1/ and we.2;2/ .
(3) Detect wo.1/ , w.2/ .3/ .3/
o , and wo separately from sw to obtain the extracted water-
.3;1/ .3;2/ .3;3/
marks we , we , and we .
3
In fact, the noises are quite loud already, as proved by the ODGs.
(4) Detect w.1/ .2/ .3/ .4/ .4/

o , wo , wo , and wo separately from sw to obtain the extracted
.4;1/ .4;2/ .4;3/ .4;4/
watermarks we , we , we , and we .
Table 5.10 shows the results of the averaging collusion attack on Bass.wav,
Gspi.wav, Harp.wav, and Pop.wav. For Bass.wav, the coded-image watermarks used
are as w.1/
o , as w.2/o , as w.3/
o , and as w.4/
o .
.1/
For Gspi.wav, the coded-image watermarks used are as wo , as w.2/
o ,
as w.3/
o , and as w .4/
o . For Harp.wav, the coded-image watermarks
used are as w.1/
o , as w.2/
o , as w.3/
o , and as w.4/
o . For Pop.wav,
.1/
the coded-image watermarks used are as wo , as w.2/
o ,
as w.3/ o , and as w .4/
o .
To evaluate the perceptual quality of the watermarked signals (including sw.1/
sw.4/ and sw.2/ sw.4/ ), their SNRs and ODGs relative to the host signal are calculated.
Also, the BERs of the extracted watermarks we are calculated to denote the detection
rate. Since w.1;1/ e can always be detected without bit errors, the results of w.1;1/ e are
omitted in the table.
It is observed from Table 5.10 that for a given host signal, the SNRs and ODGs of
sw.n/ are generally higher than sw.n/ . This means that the average watermarked signals
generally have better perceptual quality than the single watermarked signal.
Take Bass.wav as an example. On one hand, the SNRs of sw.2/ , sw.3/ , and sw.4/
are 36.75 dB, 37.54 dB, and 38.65 dB, respectively. These values are higher than
the SNRs of sw.1/ , sw.2/ , sw.3/ , and sw.4/ , which are 33.45 dB, 33.43 dB, 33.42 dB, and
33.43 dB, respectively. On the other hand, the ODGs of sw.2/ , sw.3/ and sw.4/ are 1:064,
0:644, and 0:504, respectively. These values are higher than the ODGs of sw.1/ ,
sw.2/ , sw.3/ , and sw.4/ , which are 1:761, 1:753, 1:829, and 2:001, respectively.
Furthermore, it can seen that the SNRs and ODGs of sw.2/ , sw.3/ , and sw.4/ increase
gradually. This shows that when more copies of the watermarked signals are used
in the averaging, the resulting average watermarked signal has a better perceptual
quality.
These observations result from the operation of averaging, which weakens the
effect of individual watermarks but increases the effect of the host signal. In this
way, as more copies are averaged, the average watermarked signal is more similar
to the host signal.
Meanwhile, being averaged over more copies, it become more difficult to extract
each individual watermark from the average watermarked signals. Remember that
w.;1/
e (including w.1;1/ e , w.2;1/
e , w.3;1/
e , and w.4;1/
e ) is the distorted w.1/o extracted
.1/ .2/ .3/ .4/
respectively from sw , sw , sw , and sw . It can be seen from Table 5.10 that for
a given host signal, the BERs of w.2;1/ e , w.3;1/
e , and w.4;1/
e are usually increasing. For
example, the BERs of we , we , we , and w.4;1/
.1;1/ .2;1/ .3;1/
e for Pop.wav are 0 %, 0.36 %,
4.64 %, and 6.79 %, respectively.
On the whole, the BERs of all the extracted watermarks are less than 9 %. This
indicates that the proposed scheme is quite robust against the averaging collusion
attack.
Table 5.10 Results of the averaging collusion attack

(a) Single watermarked signal
w.1/
o
Sw.1/ SNR /dB 33.45 32.05 24.48 30.41
ODG 1.761 1.757 0.531 1.179
w.2/
o
Sw.2/ SNR /dB 33.43 32.07 24.45 30.53
ODG 1.753 1.515 0.574 1.018
w.3/
o
Sw.3/ SNR /dB 33.42 32.03 24.46 30.42
ODG 1.829 1.496 0.663 1.079
w.4/
o
Sw.4/ SNR /dB 33.43 32.04 24.46 30.51
ODG 2.001 1.762 0.559 1.084
(b) Average watermarked signal
Sw.2/ SNR /dB 36.75 34.90 27.42 33.32
ODG 1.064 1.424 0.162 0.453
Sw.3/ SNR /dB 37.54 36.50 29.10 34.56
ODG 0.644 1.128 0.087 0.283
Sw.4/ SNR /dB 38.65 37.89 30.04 35.31
ODG 0.504 1.046 0.010 0.180
w.;1/
e BER: % 0 2.86 0.71 0.36
w.2;1/
e
BER: % 2.86 4.29 2.86 4.64
w.3;1/
e
BER: % 5.71 8.10 5.00 6.79
w.4;1/
e
w.;2/
e BER: % 0.32 0.57 0 0.41
w.2;2/
e
BER: % 2.22 0 0.71 4.90
w.3;2/
e
BER: % 6.67 4.57 6.43 8.16
w.4;2/
e
w.;3/
e BER: % 4.13 1.71 0.71 3.21
w.3;3/
e
BER: % 4.13 2.86 5.00 8.93
w.4;3/
e
w.;4/
e BER: % 7.30 5.24 8.57 4.49
w.4;4/
e
5.3.3.3 Test Under Multiple Watermarking
Multiple watermarking is another challenging statistical attack to audio water-

marking schemes. However, this attack has received very little attention in most
robustness tests.
As described in Sect. 3.1.3.2, multiple watermarking is to sequentially embed n

different watermarks w.1/ .2/ .n/
o ; wo ; : : : ; wo in the following way:
(
sw.1/ D Embedding so ; w.1/
o
(5.11)
sw.i/ D Embedding sw.i1/ ; w.i/
o ; 2i n
.i;j /
In the detection, i watermarks we are extracted from the watermarked signal
sw.i/individually:

w.i;j
e
/
D Detection sw.i/ ; 1 i n and 1 j i (5.12)
Similar to the averaging collusion attack, embedding multiple watermarks would

also weaken the effect of individual watermarks.
It is worth mentioning that sw.i/ in Eqs. (5.9) and (5.11) differs in the meaning,
.i;j /
similar to we in Eqs. (5.10) and (5.12).
Recall that the same or different audio watermarking techniques can be used in
multiple watermarking. For example, the first watermark w.1/o is embedded using our
proposed algorithm. The second watermark can be embedded using the proposed
algorithm again or using the echo hiding watermarking. Note that the corresponding
detection algorithm must be employed to extract the watermark. Therefore, two
types of multiple watermarking experiments are performed, i.e., multiple self-
watermarking and inter-watermarking.
• Multiple self-watermarking
In multiple self-watermarking, the host signal so is sequentially watermarked n
times by one method.
Specifically, we consider that so is sequentially watermarked four times (n D 4)
by the proposed method. The procedure for multiple self-watermarking is described
as follows:
(1) Embed w.1/ .1/ .1/ .1/
o into so to generate sw and then detect wo from sw to obtain the
.1;1/
extracted watermark we .
(2) Embed w.2/ .1/ .2/
o into sw to generate sw and then separately detect wo and wo
.1/ .2/
.2/ .2;1/ .2;2/
from sw to obtain the extracted watermarks we and we .
(3) Embed w.3/ .2/ .3/ .1/
o into sw to generate sw and then separately detect wo , wo , and
.2/
.3/ .3/ .3;1/ .3;2/ .3;3/
wo from sw to obtain the extracted watermarks we , we , and we .
(4) Embed w.4/ .3/ .4/ .1/
o into sw to generate sw and then separately detect wo , wo , wo ,
.2/ .3/
.4/ .4/ .4;1/ .4;2/ .4;3/
and wo from sw to obtain the extracted watermarks we , we , we , and
w.4;4/
e .
Table 5.11 shows the results of multiple self-watermarking on Bass.wav,
Gspi.wav, Harp.wav, and Pop.wav. For Bass.wav, the coded-image watermarks
used are as w.1/
o , as w.2/
o , as w.3/
o , and
.4/
as wo . For Gspi.wav, the coded-image watermarks used are as w.1/
o , as
Table 5.11 Results of multiple self-watermarking

w.1/
o
Sw.1/ SNR /dB 33.36 32.04 24.47 30.46
ODG 2.027 1.516 0.475 1.104
w.2/
o
Sw.2/ SNR /dB 29.79 28.49 20.85 26.66
ODG 3.131 2.672 1.702 1.959
w.3/
o
Sw.3/ SNR /dB 27.93 26.55 18.05 23.79
ODG 3.391 3.291 2.808 2.756
w.4/
o
Sw.4/ SNR /dB 26.24 25.18 16.08 21.86
ODG 3.591 3.564 3.339 3.240
w.;1/
e BER: % 0 0 0 0
w.1;1/
e
BER: % 0.29 2.86 0 0.36
w.2;1/
e
BER: % 1.43 7.43 2.86 1.79
w.3;1/
e
BER: % 4.00 7.43 2.86 2.14
w.4;1/
e
w.;2/
e BER: % 0 0 0 0
w.2;2/
e
BER: % 0.32 0 0 1.07
w.3;2/
e
BER: % 3.17 9.52 0 1.43
w.4;2/
e
w.;3/
e BER: % 0 0 0 0
w.3;3/
e
BER: % 0 1.14 0 2.14
w.4;3/
e
w.;4/
e BER: % 0 0 0 0
w.4;4/
e
w.2/
o , as w.3/ o , and as w.4/
o . For Harp.wav, the coded-image watermarks
.1/
used are as wo , as w.2/
o , as w.3/
o , and as w.4/
o . For Pop.wav, the
.1/
coded-image watermarks used are as wo , as w.2/
o , as
.3/ .4/
wo , and as wo .
For evaluation purposes, we calculate the SNRs and ODGs of the watermarked
signals (including sw.1/ sw.4/ ) relative to the host signal, as well as the BERs of the
extracted watermarks.
It is observed from Table 5.11 that for a given host signal, the SNRs and ODGs
of sw.1/ , sw.2/ , sw.3/ , and sw.4/ decrease gradually. This means that the perceptual quality
gets worse if the signal is watermarked more times. Take Bass.wav as an example.
The decreasing SNRs of sw.1/ , sw.2/ , sw.3/ , and sw.4/ are 33.36 dB, 29.79 dB, 27.93 dB,
and 26.24 dB, respectively. Also, the decreasing ODGs of sw.1/ , sw.2/ , sw.3/ , and sw.4/ are
2:027, 3:131, 3:391, and 3:591, respectively. The reason is that more samples
of the signal are modified during the embedding of more watermarks.
Meanwhile, because of being watermarked more times, each individual water-
mark becomes more difficult to extract from the multiple watermarked signals.
Remember that w.;1/
e (including we.1;1/ , we.2;1/ , we.3;1/ , and we.4;1/ ) is the distorted wo.1/
extracted respectively from sw.1/ , sw.2/ , sw.3/ , and sw.4/ . It can be seen from Table 5.11 that
for a given host signal, the BERs of we.2;1/ , we.3;1/ , and we.4;1/ are usually increasing.
For example, the BERs of we.1;1/ , we.2;1/ , w.3;1/ e , and we.4;1/ for Pop.wav are 0 %,
0.36 %, 1.79 %, and 2.14 %, respectively.
On the whole, the BERs of all the extracted watermarks are less than 10 %.
This indicates that the proposed scheme is quite robust against multiple self-
watermarking.
• Inter-watermarking
In inter-watermarking, the watermarked signal is separately re-watermarked by
other audio watermarking techniques.
In addition to the proposed method (“Proposed”), four watermarking techniques
in Chap. 3 are also considered, i.e., cepstrum domain watermarking (“Cepstrum”),
wavelet domain watermarking (“Wavelet”), echo hiding with kernel 3 (“Echo”), and
histogram-based watermarking (“Histogram”). Note that least significant bit (LSB)
modification, phase coding, and spread spectrum (SS) watermarking are not taken
into consideration, since these three watermarking techniques cannot preserve the
perceptual quality of the watermarked signals.
During the process of inter-watermarking, cepstrum domain watermarking,
wavelet domain watermarking, echo hiding, and histogram-based watermarking use
the parameters as specified in Figs. 3.8, 3.10, 3.13, and 3.14, respectively. The
results in Chap. 3 show that these parameters provide the best performance for each
watermarking technique. Also, instead of the coded-image watermarks, the PRSs
are directly embedded at full capacity. Moreover, to avoid perceived noise caused
by watermarking the silence, the PRSs are embedded into the watermarking regions
selected by the proposed method.
To make a comparison of robustness against inter-watermarking between the
proposed method and the other audio watermarking techniques, two experiments are
carried out. The first experiment is to evaluate the ability of the proposed method
resistance to inter-watermarking. The second experiment is to evaluate the ability of
the other watermarking techniques resistance to inter-watermarking.
¾ In Experiment I, the watermarked signal generated by the proposed method is
separately re-watermarked by the considered watermarking techniques.
The procedure for Experiment I is described as follows.
(1) The proposed method embeds wo into so to generate sw and then detects wo
from sw to obtain the extracted watermark we .
(2) sw is re-watermarked by the proposed method (i.e., self-watermarking) and

prop prop
sw is generated. Then, the proposed method detects wo from sw to obtain
prop
the extracted watermark we .
ceps
(3) sw is re-watermarked by cepstrum domain watermarking and sw is generated.
ceps
Then, the proposed method detects wo from sw to obtain the extracted
ceps
watermark we .
(4) sw is re-watermarked by wavelet domain watermarking and swwave is generated.
Then, the proposed method detects wo from swwave to obtain the extracted
watermark wwave
e .
(5) sw is re-watermarked by echo hiding and swecho is generated. Then, the proposed
method detects wo from swecho to obtain the extracted watermark wecho
e .
(6) sw is re-watermarked by histogram-based watermarking and swhist is generated.
Then, the proposed method detects wo from swhist to obtain the extracted
watermark whist
e .
Table 5.12 shows the results of inter-watermarking Experiment I on Bass.wav,

Gspi.wav, Harp.wav, and Pop.wav signals. The coded-image watermarks embed-
ded are for Bass.wav, for Gspi.wav, for Harp.wav, and
for Pop.wav.
For evaluation purposes, we calculate the SNRs and ODGs of the watermarked
prop
signals (including sw and sw swhist ) relative to the host signal, as well as the
BERs of the extracted watermarks.
Table 5.12 shows that most watermarks after being inter-watermarked can be
perfectly extracted. Only when the watermarked Bass.wav, Gspi.wav, and Pop.wav
signals are re-watermarked by the proposed method or histogram-based watermark-
ing, the extracted watermarks are slightly distorted. Yet the BERs are still less than
3 %.
ceps
Moreover, for a given host signal, the ODGs of sw and swwave are comparable to
prop echo hist
that of sw . However, the ODGs of sw , sw , and sw are usually much lower. Take
ceps
Pop.wav as an example. The ODGs of sw and swwave are 1:568 and 1:161, not
prop
much lower than the ODG of sw , being 1:104. However, the ODGs of sw , swecho ,
and sw are quite different, being 1:959, 2:190, and 2:305 respectively. This
hist
indicates that cepstrum domain watermarking and wavelet domain watermarking

have less influence on the perceptual quality than other techniques, which agrees
with the results in Sects. 3.2.4 and 3.2.5.
It is also found that for a given host signal, the SNRs are not in accordance with
the ODGs. Take Pop.wav as an example. The ODGs of sw and swwave are similar,
i.e., 1:104 and 1:161. However, their SNRs are quite different, i.e., 30.46 dB
and 21.35 dB. As already mentioned in Sects. 3.2.5 and 5.2.2, such an observation
motivates us to investigate other objective quality measures in Chap. 6.
¾ In Experiment II, the watermarked signals generated by the considered
watermarking techniques are separately re-watermarked by these techniques.
The above five watermarking techniques are employed as host techniques
as well as attack techniques, i.e., the proposed method (“Proposed”), cepstrum
domain watermarking (“Cepstrum”), wavelet domain watermarking (“Wavelet”),
Table 5.12 Results of inter-watermarking Experiment I

wo
sw SNR /dB 33.36 32.04 24.47 30.46
ODG 2.027 1.516 0.475 1.104
BER: % 0 0 0 0
we
Proposed SNR /dB 29.79 28.49 20.85 26.66
prop
sw ODG 3.131 2.672 1.702 1.959
BER: % 0.29 2.86 0 0.36
prop
we
Cepstrum SNR /dB 20.12 20.13 13.68 15.26
ceps
sw ODG 2.400 1.651 0.583 1.568
BER: % 0 0 0 0
ceps
we
Wavelet SNR /dB 25.23 24.22 18.25 21.35
swwave ODG 2.262 1.774 0.541 1.161
BER: % 0 0 0 0
wwave
e
Echo SNR /dB 11.07 10.26 10.79 12.09
swecho ODG 2.194 2.686 2.157 2.190
BER: % 0 0 0 0
wecho
e
Histogram SNR /dB 33.06 31.85 24.42 30.00
swhi st ODG 3.613 2.970 1.465 2.305
BER: % 1.14 1.14 0 0
whi
e
st
*
Note: Symbol “*”: the column identical to the one in Table 5.13 below
echo hiding (“Echo”), and histogram-based watermarking (“Histogram”). Each host

technique embeds the watermark into the host signal to generate the watermarked
signal sw . Then, sw is separately re-watermarked by the attack techniques and we
get the re-watermarked signals accordingly. Later, the host technique detects the
watermark from each re-watermarked signal separately.
Clearly, inter-watermarking Experiment I is one special case of inter-
watermarking Experiment II, in which the proposed method is the host technique.
Without loss of generality, Bass.wav is taken as an example. Table 5.13 shows
the results of inter-watermarking Experiment II on Bass.wav. Specifically, the
shaded cases on the diagonal line refer to each technique’s self-watermarking. For
evaluation purposes, we calculate the SNRs and ODGs of the watermarked signals
relative to the host signal, as well as the BERs of the extracted watermarks.
Since inter-watermarking Experiment I is one special case of inter-watermarking
Experiment II, the column indicated by “*” in Tables 5.13 and 5.12 shares the same
results.
Table 5.13 Results of inter-watermarking Experiment II on Bass.wav signal

Host technique
Proposed Cepstrum Wavelet Echo Histogram
Watermark length Nw 350 324 324 162 40
sw SNR /dB 33.36 20.63 26.01 10.88 44.56
ODG 2.027 0.602 0.557 2.200 2.123
BER: % 0 0 0 0 0
Attack Proposed SNR /dB 29.79 20.30 25.21 10.84 32.99
technique ODG 3:131 2.162 2.260 2.190 3.598
BER: % 0.29 0 0 0 0
Cepstrum SNR /dB 20.12 19.61 19.94 10.48 20.61
ODG 2.400 0:527 0.534 2.200 2.141
BER: % 0 0
Wavelet SNR /dB 25.23 24.60 24.57 10.78 25.85
ODG 2.262 0.574 0:604 2.200 2.149
BER: % 0 0
Echo SNR /dB 11.07 10.55 10.99 7.85 11.19
ODG 2.194 2.200 2.200 2:190 2.200
BER: % 0 0 0 0.62
Histogram SNR /dB 33.06 20.62 25.93 10.87 40.05
ODG 3.613 2.031 2.095 2.200 2:938
BER: % 1.14 0 0 0.62
*
2. Symbol “*”: the column identical to the one in Table 5.12 above
From Table 5.13, it is observed that only the proposed method and echo hiding are
robust against inter-watermarking by all five watermarking techniques. The BERs
of the extracted watermarks are less than 2 %. Note that the successful detection of
echo hiding’s self-watermarking is conditional: the echo delays used by the attack
technique are different (as far away as possible) from the ones used by the host
technique.
Meanwhile, the other three techniques fail in some cases of inter-watermarking.
For example, given the watermarked signal generated by cepstrum domain water-
marking, the embedded watermark cannot survive the re-watermarking by its self-
watermarking or wavelet domain watermarking. Similarly, given the watermarked
signal generated by wavelet domain watermarking, the embedded watermark cannot
survive the re-watermarking by its self-watermarking or cepstrum domain water-
marking. By contrast, histogram-based watermarking shows the weakest resistance
to inter-watermarking. Given the watermarked signal generated by histogram-based
watermarking, the embedded watermark can merely survive the re-watermarking by
the proposed method.
In summary, the proposed audio watermarking scheme performs well throughout
the robustness test.
5.5 Data Payload and Computational Complexity 151
5.4 Security Analysis
The goal of security analysis is to evaluate the security level of the proposed audio
watermarking scheme. As discussed in Sect. 3.2, cepstrum domain watermarking
and wavelet domain watermarking that are based on statistical mean manipulation
(SMM), echo hiding, and histogram-based watermarking all suffer from security
problem to varying degrees. A theoretical analysis of watermarking security is not
the focus of this book. As introduced in Sect. 1.3.2.3, an intuitive method of security
analysis is to calculate the possible ways for embedding. If there were more possible
ways for embedding, unauthorized detection without secret keys would become
more difficult to identify and/or remove the embedded watermark.
In our experiments, each block is divided into 32 nonlinear subbands, where
28 subbands are randomly selected for embedding. In this case, Nsubband D 32
and NQ subband D 28. Accordingly, the number of possible ways due to channel
scrambling is calculated by using Eq. (4.8):
Nscrambli ng D P Nsubband ; NQ subband D N Nsubband Š

. subband NQsubband /Š
32Š (5.13)
D P .32; 28/ D 1:1 1034
.32 28/Š
Such a huge number (i.e., 1:1 1034 ) makes unauthorized detection nearly
impossible, which means that the property of the security has increased greatly.
This is just one code complexity, which can be further multiplied by the complexity
introduced by the PRNs.
5.5 Data Payload and Computational Complexity
Data payload and computational complexity are two criteria of minor consideration
in audio watermarking for copyrights protection.
5.5.1 Estimation of Data Payload
As defined in Sect. 1.3.1.4, data payload (or capacity) of one audio watermarking
scheme is the number of bits embedded into a one-second audio fraction. According
to the embedding algorithm, the data payload of the proposed audio watermarking
scheme, DPB , is expressed as follows:
2fs Nbit
DPB D bps (5.14)
N Nc Nunit
where fs is the sampling frequency of audio signal, Nbit is the number of watermark
bits embedded per block, N is the frame length, Nc is the number of frames per unit,
and Nunit is the number of units per block. Note that factor 2 in Eq. (5.14) is due to
half-overlapping between adjacent audio frames.
Furthermore, if the coded-image watermark is adopted, the data payload in terms
of letters, DPL , is
DPB
DPL D lps (5.15)
Lw
where Lw D 35 is the number of bits comprising one letter and DPL is expressed
in letter per second (lps).
Note that the data payload discussed above refers to the theoretical data payload
of one audio watermarking scheme, which solely depends on the watermark
embedder. That is, once the embedding parameters and the embedding algorithm
used by the watermark embedder are chosen, theoretical data payload is determined
subsequently.
The values of these experiment parameters determined in Sect. 5.1 are N D 512,
Nc D 4, Nunit D 10, and Nbit D 4. Moreover, all the audio test files are sampled
at 44.1 kHz, i.e., fs D 44:1 kHz. Therefore, the data payload of the scheme under
evaluation is equal to
2 44100 4
DPB D 17:2 bps (5.16)
512 4 10
17:2
DPL D 0:5 lps (5.17)
35
which are sufficient for the purpose of copyrights protection.
From Table 5.14, however, it is observed that the watermarks embedded in
e
different host signals have quite different lengths, although the same watermark
embedder is employed. Thus, the practical data payload (DP B ) is defined as
the watermark length divided by the duration of the audio signal. For example,
we calculate the practical data payload for Bass.wav, 350 bits/24.9 s = 14.1 bps;
Gspi.wav, 210 bits/19 s = 11.1 bps; Harp.wav, 140 bits/16.4 s = 8.5 bps; and
Pop.wav, 280 bits/20 s = 14 bps. The duration of all the test signals is listed in
Appendix D. By averaging these four values, the average practical data payload of
the proposed scheme is considered to be 11.9 bps.
The practical data payload of one host signal depends on the watermark embedder
as well as the selected watermarking regions of each host signal. If a host signal
contains more silences and trifle fractions, the watermarking regions are of smaller
size and hence the practical data payload is lower. Obviously, the practical data
payload cannot always exceed the theoretical data payload. As the practical data
payload is approaching the theoretical value, more samples of the host signal are
used for watermarking.
5.5 Data Payload and Computational Complexity 153
Table 5.14 Results of the computational complexity estimation

Duration of host signal (sec) 24.9 19.0 16.4 20.0
Watermark length (bits) 350 210 140 280
Embedding time (sec) 256.1 158.7 97.0 196.3
Average embedding speed (bps) 1.37 1.32 1.44 1.43
Detection time (sec) 56.7 17.3 11.9 31.2
Average detection speed (bps) 6.17 12.14 8.15 8.97
5.5.2 Estimation of Computational Complexity
As mentioned in Sect. 1.3.1.5, computational complexity is evaluated in terms of

the speed, which is further denoted as the embedding time tembedding and detection
time .tdetection / relative to the duration of the host audio signal. Moreover, average
embedding and detection speeds are also employed to indicate the rate of embedding
and detection. That is, if a total of Nw watermark bits are embedded in time tembedding ,
the average embedding speed is given by
Nw
C Cembedding D bps (5.18)
tembedding
Similarly, if Nw watermark bits are detected in time tdetection , the average detection
speed is given by
Nw
C Cdetection D bps (5.19)
tdetection
Note that the average embedding and detection speeds are expressed in the same
unit as data payload, i.e., bps.
For the test platform, all the experiments are conducted on a Pentium 4 2.4 GHz
computer with 1 GB RAM. Table 5.14 shows the results of the computational
complexity estimation on Bass.wav, Gspi.wav, Harp.wav, and Pop.wav. Note that
for the detections measured in the experiments, watermark bits are detected from
the watermarked signals without being attacked.
From Table 5.14, it is observed that although different host signals differ with
embedding time, their average embedding speeds are similar. On the other hand,
the average detection speed of different watermarked signals varies in the detection
time as well as the average detection speed. For example, the average detection
speed of the watermarked Gspi.wav signal is almost twice as fast as than that of the
watermarked Bass.wav signal. This is due to different implementation mechanisms
of the embedding and detection algorithms in MATLAB.
During the execution of the embedding algorithm, the host signal is processed
block by block to embed watermark bits. According to Fig. 4.3, block size is merely
determined by N , Nc , and Nunit , not related to the host signal. Therefore, different
host signals have the same utilization of computer memory in the embedding,
resulting in a similar average embedding speed.
However, during the execution of the detection algorithm, the watermark bits
are extracted from each watermarking region. The reason for this is that in
the cases of desynchronization attacks, the conformation of blocks is distorted.
Therefore, the magnitudes of all tiles in every watermarking region (not just in every
block) need to be provided simultaneously for block synchronization, as shown
in Sect. 4.3.1. Consequently, different host signals have a different utilization of
computer memory in the detection. Moreover, large watermarking regions demand
more computer memory, thus the average detection speed is slow and vice versa for
small watermarking regions.
On the whole, the average detection speed is much faster than the average
embedding speed, which is a desirable attribute in copyrights protection application.
5.6 Performance Comparison
Table 5.15 compares the performance of our proposed scheme (“Proposed”) with
several existing audio watermarking schemes, sorted by chronological order. The
chosen schemes were not implemented in the book. Therefore, if the result is not
reported in the publication, it is marked by symbol = in the table. Also if the
published result is obtained or interpreted in a different way, it is marked by the
symbol .
The investigation is focused on imperceptibility (“Impcpty”), robustness, and data
payload (“Payload”), since security and computational complexity were not taken
into consideration by most schemes.
• For imperceptibility evaluation, the commonly used SNR is employed as the
metric. Moreover, it usually refers to the average SNR of all the watermarked
signals, since there are a number of host audio signals adopted in every scheme.
• For the robustness test, the attacks include noise addition (“NA”), resampling
(“RS”), amplitude scaling (“AM”), low-pass filtering (“LP”), echo addition
(“ECHO”), MP3 compression (“MP3”), and PITSM (“TSM”).4 Other attacks
such as requantization, DA/AD conversion, reverberation, random samples
cropping, jittering, zeros inserting, and TPPSM are not listed in the table, because
they were either performed in a varying way or not even conducted in most
4
The attacks with symbol in Table 5.15 are described as follows. Under the “NA” category, the
schemes in [5, 7] did not specify the value of the SNR. Under the “AM” category, the schemes
in [5, 7] compressed the amplitude with a nonlinear gain function. Under the “LP” category, the
schemes in [3, 8] tested band-pass filtering only. Under the “TSM” category, the schemes in [9, 10]
implemented random stretching (at ˙4 % and ˙8 %, respectively) merely by omitting or inserting
a random number of samples, which is considered similar to random sample cropping/inserting.
Table 5.15 Performance comparison of different audio watermarking schemes
5.6 Performance Comparison
Impcpty Robustness Payload

(SNR) NA RS AM LP ECHO MP3 TSM (DP
f B)
Proposed 30.1 dB 30 dB 22.05 kHz ˙20 % 5 kHz (0.3, 200 ms) 48 kbps ˙10 % 11.9 bps
[1] 33.5 dB 5 dB 8.82 kHz = 5 kHz = 32 kbps ˙1 % 5.4 bps
[2] 43.8 dB 40 dB 16 kHz ˙20 % 7 kHz = 64 kbps ˙25 % 2 bps
[3] 20 dB 24 kHz C50 % (0.4, 100 ms) 64 kbps = <10 bps
[4] = 20 dB 22.05 kHz 5 kHz (0.5, 100 ms) 32 kbps ˙5 % 8.4 bps
[5] 29.5 dB 22.05 kHz C50 % 4 kHz (0.4, 100 ms) 32 kbps ˙10 % 4.3 bps
[6] = 36 dB 48 kHz = 4 kHz = 56 kbps ˙10 % 2.3 bps
[7] 32.4 dB 22.05 kHz 8 kHz (0.5, 100 ms) 48 kbps ˙3 % 11.8 bps
[8] = 36 dB = = (=, 100 ms) 32 kbps ˙4 % 0.5-1 bps
[9, 10] 35 dB 40 dB = = = (0.5, 100 ms) 96 kbps 2.2 bps
155
schemes.5 It is worth mentioning that we do not compare the detailed results

of the BER between the schemes, but list their tolerance for each attack that was
reported in their publications. The reason is that different host audio signals were
used in different schemes, and moreover, some schemes calculated the BERs in
e
a different way. Thus, there is no direct comparison in terms of the BER.
• For data payload estimation, the practical data payload (DP B ) instead of the
theoretical data payload is adopted for comparison. Since most schemes have
no theoretical analysis on the data payload, the actual amount of watermark bits
embedded into one host signal of certain duration is calculated as the practical
data payload. Similar to the SNRs, the practical data payloads shown in the table
are usually the average values of all the watermarked signals in each scheme.
Table 5.15 shows that these schemes have different performance characteristics.
On the average, the proposed scheme achieves the best compromise between
imperceptibility, robustness, and capacity.
• In terms of imperceptibility, the SNR of the proposed scheme is within the range
of other schemes. Also, the average ODG of the proposed scheme is 1:33,
obtained from Table 5.2. In spite of a higher SNR, the ODGs reported in the
scheme in [2] are around 1:80, not superior to the proposed scheme. Without
addressing the SNR, the scheme in [3] has an average ODG of 0:93.
• In terms of robustness, the focus is on the performance under PITSM, as well
as noise addition, low-pass filtering, and MP3 compression, since most schemes
show high resistance to resampling, amplitude scaling, and echo addition. Under
PITSM, only the proposed scheme and the schemes in [2, 5, 6] can resist
excessive distortion of up to ˙10 % or greater, and hence are chosen for further
comparison. The proposed scheme is robust against PITSM (˙10 %), noise
addition (30 dB), low-pass filtering (5 kHz), and MP3 compression (48 kbps).
Compared to the proposed scheme, the scheme in [2] is quite robust against
PITSM (˙25 %), but relatively vulnerable to noise addition (40 dB), low-pass
filtering (7 kHz), and MP3 compression (64 kbps). The scheme in [5] is slightly
more robust against Low-pass filtering (4 kHz) and MP3 compression (32 kbps);
nevertheless the SNR has no specified value to compare the robustness against
noise addition.6 The scheme in [6] is slightly more robust against low-pass
5
These unlisted attacks were undertaken in several schemes as follows. Requantization: only
the scheme in [3] tested 8-bit requantization and the detection succeeded. DA/AD conversion:
the schemes in [4, 7–10] tested DA/DA conversion and the detections succeeded. Cropping: the
schemes in [2, 4, 5] tested different cropping operations and the detections succeeded. Jittering:
the schemes in [2, 5] tested different jittering operations and the detections succeeded. TPPSM:
the scheme in [1] tested ˙1 % pitch-scaling and the detection succeeded; the scheme in [3] tested
the case that the pitch is shifted up by two semitones and the detections completely failed; the
schemes in [9, 10] implemented pitch shifting (at ˙4 % and ˙8 % respectively) merely by linear
interpolation without anti-alias filtering and the detections succeeded.
6
It was reported as “noise addition that can be heard clearly by everybody [5].”
5.7 Summary 157
filtering (4 kHz), but less against noise addition (36 dB) and MP3 compression
(56 kbps).
• In terms of the data payload, the proposed scheme has the highest practical
data payload among these schemes, i.e., 11.9 bps. In particular, this value is
much higher than data payloads of the schemes in [2, 5, 6], i.e., 2 bps, 4.3 bps,
and 2.3 bps respectively. Moreover, as shown in Eq. (5.16), the theoretical data
payload of the proposed scheme is even higher—about 17.2 bps.
5.7 Summary
In this chapter, the performance of the proposed audio watermarking scheme has
been thoroughly evaluated with respect to imperceptibility, robustness, security,
data payload, and computational complexity. Specifically, the designed performance
evaluation consists of perceptual quality assessment, robustness test, security
analysis, estimations of the data payload, and computational complexity. Without
loss of generality, the performance evaluation presented in this chapter can serve as
one comprehensive benchmark of audio watermarking algorithms.
Firstly, the subjective listening test and the objective evaluation test were
employed in the perceptual quality assessment. Specifically, the subjective listening
test includes the MUSHRA test and SDG rating, while the objective evaluation test
includes the calculation of the ODG (using PEAQ) and the SNR value. Secondly,
both basic and advanced robustness tests were carried out. Basic robustness test
includes common signal operations (e.g., noise addition, resampling, requantization,
amplitude scaling, low-pass filtering, DA/AD conversion, echo addition, reverber-
ation, and MP3 compression), desynchronization attacks (e.g., random samples
cropping, jittering, zeros inserting, PITSM, and TPPSM), and combined attacks
(e.g., Type I and Type II combined attacks). The advanced robustness test includes
StirMark for Audio, averaging collusion, and multiple watermarking (e.g., self-
watermarking and two types of inter-watermarking). Thirdly, the number of possible
embedding ways due to channel scrambling was calculated in the security analysis.
Furthermore, both theoretical and practical data payloads were calculated. Finally,
computational complexity was evaluated in terms of the embedding/detection PC
computing time as well as the average embedding/detection speed.
The experimental results show that the watermarked audio signals are perceptu-
ally transparent, robust against various attacks, and self-secured from unauthorized
detection. Also, watermarking efficiency of the proposed technique is satisfactory
with respect to the data payload and computational complexity as compared to the
other methods.
Compared with other reported schemes, the proposed scheme achieves a better
compromise between imperceptibility, robustness, and data payload. Thus, it is
concluded that the proposed audio watermarking scheme performs well for the
purpose of copyrights protection.
Chapter 6
Perceptual Evaluation Using Objective
Quality Measures
Imperceptibility is a prerequisite to the use of the watermarked audio; hence, per-

ceptual quality assessment is worthy of more attention. Objective quality measures
have been widely used in speech quality evaluation. In this chapter, we introduce
objective quality measures used for the first time in the perceptual quality evaluation
of audio watermarking.
Perceptual quality assessment in audio watermarking including subjective lis-
tening tests and objective evaluation tests is reviewed first. This is followed by
a description of the objective quality measures under investigation. Next, several
experiments are performed to explore the relations between objective quality
measures and perceptual quality in the context of different audio watermarking
techniques. Finally, some comments are made to summarize the performance of
the considered objective quality measures as the perceptual quality predictors.
6.1 Perceptual Quality Evaluation
The aim of audio watermarking is to embed an imperceptible, robust, and secure

watermark into host signals. From the viewpoint of communication theory, the
watermark is inserted into a cover signal like a kind of noise. Considering that the
process of watermarking should be perceptually transparent, the perceptual quality
of the watermarked audio signal is evaluated relative to the host audio signal.
As described in Chap. 1, there are two approaches to perceptual quality assess-
ment of audio watermarking: (1) subjective listening tests by human acoustic
perception and (2) objective evaluation tests by perception modelling or quality
measures.
• Subjective listening test
In the subjective listening tests, the listeners are asked to compare the perceptual
quality of the watermarked audio signal with the host audio signal. As stated in
160 6 Perceptual Evaluation Using Objective Quality Measures
Sect. 1.3.2.1, the ABX test and the MUSHRA test (i.e., MUlti Stimuli with Hidden
Reference and Anchors) are two commonly used methods.
In the ABX listening test (see Appendix B), the listener has to identify an
unknown sample X as being A or B, with A (the host signal) and B (the
watermarked signal) available for reference. Initially, the ABX test was designed
for the assessment of small deterioration [43]. Note that ABX tests can also
be performed as ABC/HR tests, i.e., double blind, triple stimulus, with hidden
reference [31]. Specifically, stimulus A is the host signal for reference, whereas
stimulus B and C are the host and watermarked signals in randomized order. After
listening to three stimuli, the listener is asked to decide between B and C as the
hidden reference signal, and then the remaining one is the watermarked signal.
Finally, the watermarked signal is evaluated relative to the host signal by using a
subjective difference grade (SDG), as described in Table 1.2.
The MUSHRA test (see Sect. 5.2.1) is developed for assessing intermediate
audio quality [44]. Since multiple stimuli including the hidden reference and a few
additional signals (anchors) are employed, the MUSHRA test is supposed to be
more reliable than the ABX test in the presence of slightly larger distortions.
Subjective listening tests are indispensable to perceptual quality assessment,
since the ultimate judgment is made by human perception. However, it is quite
time-consuming and cost intensive to conduct such listening tests. Moreover, the test
results are subject to test environments and the participants’ preferences. Therefore,
machine-based objective evaluations are used to provide a convenient, consistent,
and fair assessment.
• Objective evaluation test
Objective evaluation test is intended to facilitate the implementation of subjective
listening test. To achieve this goal, the results of objective evaluation should
correlate well with the SDG scores.
Currently, the commonly used objective evaluation is to assess the perceptual
quality of audio data via a stimulant ear, such as Evaluation of Audio Quality
(EAQUAL) [47], Perceptual Evaluation of Audio Quality (PEAQ) [48], and Percep-
tual Model-Quality Assessment (PEMO-Q) [49]. Basically, these methods establish
an auditory perception model to imitate the listening behavior of a human being, so
that the watermarked signal is graded relative to the host signal. The whole process
is depicted in Fig. 6.1 [31, 46]. After the watermark is embedded, the host and
watermarked signals are separately passed to a psychoacoustic model. As described
in Sect. 2.4, the psychoacoustic model calculates the internal representation of signal
features, such as the masking threshold. By comparing the internal representations
of the host and watermarked signals, the audible difference is determined. The
audible difference is the input to the cognitive model, which models the cognitive
processes in the human brain. After the audible difference is perceptually scaled
in the cognitive model, the final output is an objective difference grade (ODG). As
mentioned in Sect. 3.1.2, the specifications of ODG conform to those of SDG. To
guarantee the accuracy of evaluation, a large set of relevant test signals are required
to train and characterize such models [46].
6.2 Objective Quality Measures 161
Watermark
Host signal embedder
Watermarked signal
Psychoacoustic Psychoacoustic
model model
Internal representation Internal representation
Comparison
Audible difference
Cognitive
model
Objective Difference Grade

(ODG)
Fig. 6.1 Objective evaluation via perception modelling
Among the implemented models, PEMO-Q is the latest and most advanced
predictor of audio quality. It is reported in [49] that PEMO-Q has a higher ability to
be applicable to unknown distortions and performs better than the other techniques.
The performance of three evaluation tools will be examined in Sect. 6.3.3.
Besides perception modelling, the extent of dissimilarity between the water-
marked and host signals can be quantified by objective quality measures. Objec-
tive quality measures, such as the signal-to-noise ratio measure, the segmental
signal-to-noise ratio measure, the cepstral distortion measure, the log-likelihood
ratio measure, the Itakura–Saito distortion measure, the log-area ratio measure,
and the weighted spectral slope measure [50], are commonly used in speech
processing. They have been widely used in quality evaluation for speech enhance-
ment [136–138], speech intelligibility estimation [139], speech recognition in blind
source separation [140, 141], and noise reduction schemes [138]. We investigate
using these quality measures for objective assessments of the perceptual quality of
audio watermarking for the first time [142, 143].
6.2 Objective Quality Measures
Objective quality measures have been widely used in the quality evaluation of
speech signals [50]. This kind of measurement makes use of sound source infor-
mation and calculates the distance or distortion of the test signal with regard to
the original signal [140], which corresponds to the concept of perceptual quality
assessment in audio watermarking.
As discussed in Chaps. 3 and 5, the signal-to-noise ratio (SNR) has already been
employed to quantify the distortion that a watermark imposes on the host signal.
However, the SNR actually averages the distortions on the entire signal. Thus, it is
not an accurate indicator of perceptual quality, as indicated in Sects. 3.2.5 and 5.2.2.
Based on the results in the existing literature, six more quality measures are
selected to estimate the distance between the host and watermarked signals. Since
the impact of noise on signal quality is nonuniform, all the measures calculate the
level of distortion for each frame. As a convention, the subscripts o and w denote
the components related to host frame and the watermarked frame, respectively.
• Segmental signal-to-noise ratio (segSNR) measure
The segSNR is a variation of the SNR, obtained by averaging the SNRs of all the
frames. Referring to the formula for the SNR in Eq. (1.3), the frame-based segSNR
is calculated by [136, 138, 140]
X
N
Œgo .n/2
nD1
dsegSNR .gw ; go / D 10 log10 (6.1)
X
N
Œgw .n/ go .n/2
nD1
where go is the host frame, gw is the watermarked frame, and N is the frame length
in samples. In our experiments, N D 512, which corresponds to 11.6 ms for a
sampling rate of 44.1 kHz.
In fact, frames with segSNRs above 35 dB do not reflect human perceptual
differences; therefore, their segSNRs are generally replaced with 35 dB. Moreover,
silence frames have negative segSNRs because the signal energy is small. To prevent
getting such abnormal segSNRs, a lower threshold for the segSNR is set to be
10 dB. Thus, the segSNR values are limited in the range of [10 dB, 35 dB]
[50, 136, 138].
• Cepstral distortion measure
The cepstral distortion (CD) measure provides an estimate of cepstral distance
between the watermarked frame and the host frame. Given both cepstral coefficient
vectors cEw and cEo , CD for the first L coefficients is calculated by [140]
X
L
2
dCD cEw ; cEo D cEw .l/ cEo .l/ (6.2)
lD1
where L D 50 in our experiments.

6.2 Objective Quality Measures 163
• Log-likelihood ratio measure

The log-likelihood ratio (LLR or Itakura distance) measure is based on linear
prediction (LP) analysis. Given both LP coefficient vectors aEw and aEo , LLR measure
is defined by [50, 136, 138, 140]

aEw Ro aEwT
dLLR aEw ; aEo D log10 (6.3)
aEo Ro aEoT
where Ro is the autocorrelation matrix and ./T refers to the transpose of a matrix.
• Itakura–Saito distortion measure
The Itakura–Saito (IS) distortion measure is slightly different from the LLR measure
and defined by [50, 138, 140]
2 2
aEw Ro aEwT
dIS aEm ; aEo D o
C log10 w
1 (6.4)
2
w aEo Ro aEoT 2
o
where o2 and w2 are all-pole gains for the host and watermarked frames, respec-
tively.
It was mentioned in [140] that LLR and IS measures perform well as predictors
of the recognition rate for the signals with additive noise in continuous speech
recognition systems.
• Log-area ratio measure
The log-area ratio (LAR) measure is also based on LP analysis in that it depends on
LP reflection coefficients [136–138, 140]:
ˇ ˇ
ˇ X 2 ˇ1=2
ˇ1 P 1 C rEo .p/ 1 C rEw .p/ ˇˇ
dLAR rEw ; rEo D ˇˇ log10 log10 (6.5)
ˇ P pD1 1 rEo .p/ 1 rEw .p/ ˇˇ
where P is the order of LP analysis and P D 10 in our experiments. rEo and rEw are the
LP reflection coefficient vectors of the host and watermarked frames, respectively.
Since the reflection coefficients are closely related to power spectra, the LAR
measure is able to estimate the differences between the logarithms of the spectra of
the host and watermarked signals efficiently [137]. In [136, 137, 140], it has been
observed that the LAR is the best measure in some cases.
• Weighted spectral slope measure
The Weighted spectral slope (WSS) measure is based on an auditory model, in
which 36 overlapping filters of progressively larger bandwidth are used to estimate
the smoothed short-time spectra [136]. Then, the weighted difference between the
spectral slopes .SL/ in each band are calculated [139].
According to [50, 136, 138, 140], the WSS measure in decibels is formulated as
X
36
dW SS D Kspl .Ko Kw / C wa .k/ ŒSLo .k/ SLw .k/2 (6.6)
kD1
where Ko and Kw are related to the overall sound pressure level and Kspl is a
parameter that can be varied to increase overall performance. In our experiments,
Kspl D 0 is used as in [140] and the weight wa depends on the formant locations
[136]. As concluded in [50, 140], the WSS measure might outperform other
measures because it employs the auditory model.
Note that for each objective quality measure (except segSNR), its overall quality
score is obtained by using the m95 % mean to reduce the number of outliers. The
m95 % mean of each quality measure is calculated in the following way. First, the
value of the quality measure is calculated for each frame. Then the values of the
quality measure for all the frames are sorted in an ascending order. The m95 % mean
is the average of the first 95 % values of each quality measure [136, 138].
6.3 Experiments and Discussion
In this section, objective quality measures are evaluated to estimate their capabilities
for predicting the perceptual quality of the watermarked audio signals. This is
achieved by performing correlation analysis between the SDGs and the values of
objective quality measures.
The audio signals used are taken from the test set prepared in Sect. 3.1.1, 17
pieces of audio signals (A1 A17 ) in total. However, PEMO-Q is a commercial
software tool and its demo version is strictly limited to signal lengths up to 4 s.
Therefore, we always use a 4 s length from the beginning of each original audio test
signal and then utilize them for the experiments here. Moreover, all the simulations
are also conducted on a Pentium 4 2.4 GHz computer with 1 GB RAM under
Windows XP operating system.
6.3.1 Audio Watermarking Techniques Default Settings
The performance of objective quality measures are fully investigated under different
audio watermarking techniques, such as the proposed scheme in Chap. 4 along with
cepstrum domain watermarking, wavelet domain watermarking, echo hiding, and
histogram-based watermarking in Chap. 3.
In the experiments, each technique is employed to implement the process of
watermarking separately. The imperceptibility of the watermarked signal is con-
trolled by the watermark strength or a similar factor. Without loss of generality, all
6.3 Experiments and Discussion 165
five watermarking techniques directly embed the pseudorandom number sequence

(PRS) at full capacity. Based on the results in Chaps. 3 and 5, each watermarking
technique uses the following parameters which have provided the best perfor-
mance:
(1) Proposed audio watermarking scheme: frame length N D 512, the number
of units per block Nunit D 10, the number of watermark bits embedded in
one block Nbit D 4, the number of slots for embedding one watermark bit
NB D 30, the number of selected subbands NQ subband D 28, and watermark
length Nw D 64. In the experiments, every host signal is watermarked twenty
times with a watermark strength of ˛w D 10; 20; 30; : : : ; 200.
(2) Cepstrum domain watermarking: frame length N D 2048, repetition coding
nr D 3, and watermark length Nw D 57. In the experiments, every host signal
is watermarked eleven times with a watermark strength of ˛w D 1 103 ; 1:2
103 ; 1:4 103 ; : : : ; 3 103 .
(3) Wavelet domain watermarking: frame length N D 2048, repetition coding
nr D 3, and watermark length Nw D 57. In the experiments, every
host signal is watermarked ten times with a watermark strength of ˛w D
0:01; 0:02; 0:03; : : : ; 0:1.
(4) Echo hiding (kernel 3): frame length N D 4096, repetition coding nr D 3,
and watermark length Nw D 28. In the experiments, every host signal is water-
marked eleven times with an echo amplitude of ˛ D 0:1; 0:12; 0:14; : : : ; 0:3.
(5) Histogram-based watermarking: the embedding strength Eh D 1:4 and water-
mark length Nw D 10. In the experiments, every host signal is watermarked
eleven times with an embedding range of D 2; 2:05; 2:1; : : : ; 2:5.
Note that for each watermarking technique, every host signal has N˛ watermarked
signals: N˛ D 20 for the proposed scheme, N˛ D 11 for cepstrum domain
watermarking, N˛ D 10 for wavelet domain watermarking, N˛ D 11 for echo
hiding, and N˛ D 11 for histogram-based watermarking. The notation N˛ will be
used for correlation analysis to be undertaken in Sect. 6.3.4.
6.3.2 Subjective Listening Tests
Similar to the previous subjective listening tests, ten trained listeners participated
in the tests that were performed in an isolated chamber. Also, all the stimuli were
presented through a high-fidelity headphone.
During the tests, the participants were asked to evaluate the perceptual quality
of the watermarked signal relative to its host signal and subsequently provide a
SDG. In view of the difficulties in the real listening tests, only the proposed audio
watermarking scheme was considered. Moreover, for each host signal, the human
subjects were not required to evaluate all the twenty watermarked signals (i.e.,
˛w D 10; 20; 30; : : : ; 200), but just five of them with ˛w D 40; 80; 120; 160; 200.
In addition, the host signal was continuously included as a watermarked signal
with ˛w D 0. This is because subjects are apt to make incorrect judgments in a

situation where audible differences between test signals are too subtle to perceive.
Therefore, for each host signal, there were six watermarked signals with ˛w D
0; 40; 80; 120; 160; 200.
In this way, every listener needed to participate in 17 separate tests on Ai , i D
1; 2; : : : ; 17. For each Ai , the six watermarked signals of Ai are denoted by Aij 0 ,
0
j D 1; 2; : : : ; 6. Since there were ten subjects participating in the
tests, the SDG
0
score of Aij 0 provided by the k-th subject is denoted by GSDG i; j ; k , where
k D 1; 2; : : : ; 10. Then, the average SDG for host signal Aij 0 is calculated as
0
1 X
K 0

GQ SDG i; j D GSDG i; j ; k (6.7)
K
kD1
where K D 10.
For simplicity of expression, the
n average
SDGso for0 each host signal Ai is denoted
Q Q Q 0
as GSDG .i /, where GSDG .i / D GSDG i; j , j D 1; 2; : : : ; 6.
6.3.3 Objective Evaluation Tests
Objective evaluation tests comprise two stages: investigation of the evaluation tools
and calculation of the values of the quality measures.
• Evaluation tool analysis
In the first stage, we investigate the effectiveness of three evaluation tools using
perception modelling, namely PEMO-Q [49], EAQUAL [47], and PEAQ [48]. The
aim is to find the best quasi-subjective predictor of audio quality that would best
conform to the SDG. Its ODGs will be adopted subsequently as quasi-SDGs for
correlation analysis in the next section. The reason of using quasi-SDGs rather than
SDGs is that it would be inaccurate to perform a correlation with an insufficient
amount of the average SDGs.
To this purpose, all the watermarked signals of each host signal are evaluated
separately using three evaluation tools.
Take the proposed audio watermarking scheme as an example. Each host signal
has twenty watermarked signals with ˛w D 10; 20; 30; : : : ; 200. Moreover, the
host signal is also included to correspond with its GQ SDG obtained above. So
for each host signal Ai , there are twenty-one watermarked signals with ˛w D
0; 10; 20; 30; : : : ; 200. After being evaluated by three
n tools, eachohost signal Ai
receives three kinds of ODGs, i.e., GODG1 .i / D GODG1 i; jO by PEMO-Q,
n o n o
GODG2 .i / D GODG2 i; jO by EAQUAL, and GODG3 .i / D GODG3 i; jO
by PEAQ, where jO D 1; 2; : : : ; 21.

G SDG
Bass.wav GODG1
0 GODG2
−2 GODG3
−4
Gspi.wav
Difference Grade (SDG & ODG)
−2
−4
Harp.wav
0
−2
−4
Pop.wav
0
−2
−4
0 20 40 60 80 100 120 140 160 180 200
Watermark strength
Fig. 6.2 Evaluation of PEMO-Q, PEAQ, and EAQUAL
Figure 6.2 shows the average SDGs and the ODGs for the watermarked Bass.wav,
Gspi.wav, Harp.wav, and Pop.wav signals, respectively.
Note that a few ODGs in Fig. 6.2 are slightly positive, as with some values with
small watermark strengths. As mentioned in Sect. 3.2.1, such cases are interpreted
as distortions that are mostly inaudible for humans.
According to Fig. 6.2, for each host signal, its GODG1 are closer to GQ SDG
than GODG2 and GODG3 . It means that PEMO-Q provides a better correspondence
between subjective and objective difference grades. Therefore GODG1 .i / or simpli-
fied as GQ .i / are used for correlation analysis.
• Quality measures calculation
In the second stage, we calculate the values of the quality measures between the
host signal and all its watermarked signals. The selected objective quality measures
include the SNR, segSNR, CD, LLR, IS, LAR, and WSS measures, denoted by
r D 1; 2; : : : ; 7, respectively. Consequently, each host signal Ai has the values of
seven quality measures, denoted by Or .i /, r D 1; 2; : : : ; 7.
Table 6.1 Comparison of the total computation time (s)

PEMO-Q SNR segSNR CD LLR IS LAR WSS
55.48 0.012 0.051 3.78 2.19 2.20 1.82 2.24
Different from the tests in the first stage, the watermarked signals with ˛w D 0
are not included in quality measures calculation. Otherwise, the values of the SNR
and the segSNR are infinite, while the values of other measures are always equal to
zero.
Take the proposed audio watermarking scheme as an example. Each host signal
has twenty watermarked signals with ˛w D 10; 20; 30; : : : ; 200. Thus, for each
host signal Ai , the values of seven quality measures are calculated separately, i.e.,
O1 .i / D fO1 .i; j /g for the SNR measure, O2 .i / D fO2 .i; j /g for the segSNR
measure, O3 .i / D fO3 .i; j /g for the CD measure, O4 .i / D fO4 .i; j /g for the
LLR measure, O5 .i / D fO5 .i; j /g for the IS measure, O6 .i / D fO6 .i; j /g
for the LAR measure, and O7 .i / D fO7 .i; j /g for the WSS measure, where
j D 1; 2; : : : ; 20.
As the watermarked signal with ˛w D 0 is excluded in the above calculations,
the length of Or .i / is always one less than that of GQ .i /. To conduct a correlation
analysis, the first value of GQ .i / that corresponds to the watermarked signal with
˛w D 0 is discarded, so that GQ .i / has the same length as Or .i /.
Then we repeat the above procedure of calculating the values of quality measures
for different audio watermarking techniques.
In summary, for a given watermarking technique, each host signal Ai has N˛
watermarked signals, as introduced in Sect. 6.3.1. Based ˚ on these N˛ watermarked
signals, each Ai receives the quasi-SDGs GQ .i / D GQ .i; j / and the values
of seven quality measures Or .i / D fOr .i; j /g, r D 1; 2; : : : ; 7, where j D
1; 2; : : : ; N˛ .
Note that computation time is also one of our concerns. Table 6.1 lists the
computation time of quality measures on one watermarked Bass.wav signal with
˛w D 60 in the proposed audio watermarking scheme. PEMO-Q took around 55 s
to complete the evaluation of one watermarked signal with the default settings in
[49]. In comparison, all quality measures finished in less than 4 s, much faster
than PEMO-Q. Particularly the SNR and segSNR measures took the least time,
less than 0.1 s. Also, the computation time of the LAR, LLR, and IS measures are
not more than 2.2 s. The measured response times are based on using Pentium 4 PC
empowered by 2.4 GHz CPU and 1 GB RAM running under Windows XP operating
system.
6.3.4 Performance Evaluation Using Correlation Analysis
To evaluate the performance of objective quality measures, Pearson correlation

coefficient is calculated between the values of each quality measure, Or , and its
quasi-SDGs, GQ . Commonly, is defined by [138–140]
X
Nm

Or .n/ O r GQ .n/ G Q
nD1
D ( ) 1=2 ( ) 1=2 (6.8)
X
Nm
2 X
Nm
2
Or .n/ O r GQ .n/ G Q
nD1 nD1
where Nm is the length of Or and GQ . O r and G Q are the means of Or and GQ ,

respectively.
Note that a correlation coefficient is a number between 1 and 1. If the coefficient
is closer to 1 (positive correlation) or 1 (inverse correlation), it indicates that the
values of quality measure are in higher correlation with the quasi-SDGs. If the
coefficient is closer to 0, it indicates that there is less correlation between the values
of quality measure and the quasi-SDGs.
For each audio watermarking technique, two types of correlation analyses are
conducted [138]. Recall that the indices i , j , r are for indicating the host signal,
watermark strength, and quality measure, respectively.
In the first analysis, the correlation is separately performed on each host signal.
Given a host signal Ai , its individual correlation coefficient with the rth objective
quality measure .i; r/, i D 1; 2; : : : ; 17 and r D 1; 2; : : : ; 7 is calculated by [143]
X
N˛

Or .i; j / O r .i / GQ .i; j / G Q .i /
j D1
.i; r/ D 8 91=2 8 91=2
<XN˛
= <X N˛
=
2 2
Or .i; j / O r .i / GQ .i; j / G Q .i /
: ; : ;
j D1 j D1
(6.9)
X
N˛ X
N˛
where O r .i / D 1
N˛
Or .i; j / and G Q .i / D 1
N˛
GQ .i; j /.
j D1 j D1
The average correlation coefficient of each quality measure 1 .r/, r D
1; 2; : : : ; 7 is calculated by
1 X
Nh
1 .r/ D .i; r/ (6.10)
Nh iD1
where Nh D 17 is the number of host signals used in the experiments.

In the second analysis, the correlation is directly performed on all the host signals.
The overall correlation coefficient of the rth objective quality measure 2 .r/, r D
1; 2; : : : ; 7 is calculated by [143]
X
Nh X
N˛

Or .i; j / O r GQ .i; j / G Q
iD1 j D1
2 .r/ D 8 91=2 8 91=2
<XNh X
N˛
= <X Nh X
N˛
=
2 2
Or .i; j / O r GQ .i; j / G Q
: ; : ;
iD1 j D1 iD1 j D1
(6.11)
X
Nh X
N˛ X
Nh X
N˛
where O r D 1
Nh Na
Or .i; j / and G Q D 1
Nh Na
GQ .i; j /.
iD1 j D1 iD1 j D1
Note that the average correlation coefficient is widely used in studying objective
quality measures [136–141]. The overall correlation coefficient is more desirable in
some applications, but considered to be rather stringent [138, 144].
Tables 6.2, 6.3, 6.4, 6.5, and 6.6 show the Pearson correlation coefficients
under different audio watermarking techniques. The results include the individual
correlation coefficients .i; r/, the average correlation coefficients (absolute value)
j1 .r/j, and the overall correlation coefficients (absolute value) j2 .r/j, where
i D 1; 2; : : : ; 17 for denoting the host signal Ai and r D 1; 2; : : : ; 7 for denoting
the quality measure. In each table, the highest j1 .r/j and j2 .r/j (i.e., the absolute
value closer to 1) are shaded and the second highest ones are in bold.
Some observations can be obtained from the Pearson correlation coefficients
[143].
• The overall correlation coefficients j2 .r/j are generally lower than the average
correlation coefficients j1 .r/j under different audio watermarking techniques.
Take the proposed audio watermarking scheme in Table 6.2 as an example.
j1 .r/j, r D 1; 2; : : : ; 7 are equal to 0.92, 0.87, 0.92, 0.95, 0.85, 0.95, and 0.94,
respectively, not less than 0.85. However, j2 .r/j, r D 1; 2; : : : ; 7 are equal to
0.38, 0.30, 0.26, 0.58, 0.27, 0.64, and 0.59, respectively, not more than 0.64.
This is because the functional relationship between objective quality measure
and the quasi-SDGs varies across different types of audio signals, i.e., vocal,
percussive instrument, tonal instrument, and music. Even in the same category,
different instruments or different genres of music are most likely to exhibit
different time–frequency characteristics. Consequently, the overall correlation
coefficients j2 .r/j are less than average correlation coefficients j1 .r/j in most
cases of audio watermarking techniques, whereas exceptions exist, due to the
intricacy of different techniques.
If audio signals have similar properties, the overall correlation coefficients
become better. For instance, host audio signals A1 (Soprano.wav), A2 (Bass.wav),
and A3 (Quartet.wav) all belong to the vocal category and also have the same
lyrics. Then the overall correlation coefficients over A1 , A2 , and A3 can be
calculated by eq. (6.11), where Nh D 3. Figure 6.3 shows the results of 2 .r/
Table 6.2 Pearson correlation coefficients under our proposed scheme

SNR segSNR CD LLR IS LAR WSS
rD1 rD2 rD3 rD4 rD5 rD6 rD7
Individual Ai D1 0.97 0.86 0.97 0.98 0.86 0.98 0.93
coeff. Ai D2 0.98 0.95 0.93 0.99 0.79 0.99 0.96
.i; r/ Ai D3 0.92 0.99 0.99 0.94 0.93 0.95 0.98
Ai D4 0.93 0.93 0.95 0.96 0.89 0.98 0.96
Ai D5 0.89 0.95 0.98 0.96 0.97 0.93 0.99
Ai D6 0.72 0.36 0.47 0.71 0.33 0.72 0.57
Ai D7 0.98 0.91 0.88 0.99 0.78 0.98 0.90
Ai D8 0.91 0.97 0.97 0.99 0.94 0.96 0.96
Ai D9 0.93 0.98 0.97 0.98 0.90 0.97 0.93
Ai D10 0.97 0.64 0.90 0.98 0.74 0.98 0.96
Ai D11 0.93 0.98 0.98 0.96 0.94 0.93 0.97
Ai D12 0.97 0.94 0.94 0.97 0.82 0.96 0.98
Ai D13 0.92 0.50 0.81 0.97 0.60 0.94 0.89
Ai D14 0.87 0.98 0.97 0.85 0.98 0.94 0.99
Ai D15 0.91 0.99 0.99 0.99 0.94 0.98 0.98
Ai D16 0.89 0.98 0.98 0.97 0.98 0.94 0.99
Ai D17 0.88 0.96 0.97 0.98 0.97 0.94 0.99
Average coeff. j1 .r/j 0.92 0.87 0.92 0.95 0.85 0.95 0.94
Overall coeff. j2 .r/j 0.38 0.30 0.26 0.58 0.27 0.64 0.59
Table 6.3 Pearson correlation coefficients under cepstrum domain watermarking

Individual Ai D1 0:82 0:78 0:62 0:93 0:87 0:92 0:90
coeff. Ai D2 0:84 0:83 0:88 0:99 0:99 0:99 0:81
.i; r/ Ai D3 0:80 0:79 0:87 1:00 0:99 0:98 0:79
Ai D4 0:94 0:95 0:85 0:94 0:56 0:96 0:97
Ai D5 0:81 0:68 0:87 0:95 0:99 0:89 0:76
Ai D6 0:72 0:92 0:82 0:81 0:43 0:89 0:83
Ai D7 0:96 0:96 0:93 0:91 0:74 0:98 0:98
Ai D8 0:74 0:66 0:84 0:98 0:99 0:90 0:86
Ai D9 0:79 0:70 0:80 1:00 1:00 0:92 0:91
Ai D10 0:92 0:90 0:89 0:90 0:84 0:98 0:97
Ai D11 0:46 0:39 0:38 0:78 0:90 0:67 0:64
Ai D12 0:96 0:95 0:98 0:91 0:81 0:98 0:98
Ai D13 0:91 0:94 0:57 0:77 0:41 0:85 0:71
Ai D14 0:72 0:68 0:82 0:99 0:99 0:93 0:76
Ai D15 0:81 0:77 0:77 1:00 1:00 0:95 0:85
Ai D16 0:81 0:77 0:82 1:00 1:00 0:96 0:84
Ai D17 0:23 0:31 0:20 0:12 0:14 0:11 0:26
Average coeff. j1 .r/j 0:75 0:73 0:73 0.87 0:79 0.86 0:78
Overall coeff. j2 .r/j 0:36 0:39 0:17 0.72 0:37 0.70 0:49
Table 6.4 Pearson correlation coefficients under wavelet domain watermarking

Individual Ai D1 0:86 0:84 0:77 1:00 1:00 0:99 0:96
coeff. Ai D2 0:86 0:84 0:88 0:99 1:00 0:99 0:89
.i; r/ Ai D3 0:93 0:91 0:84 0:97 0:98 1:00 0:93
Ai D4 0:96 0:97 0:79 0:94 0:71 0:96 0:93
Ai D5 0:85 0:79 0:90 0:99 0:99 0:96 0:82
Ai D6 0:74 0:91 0:90 0:82 0:60 0:90 0:83
Ai D7 0:98 0:97 0:90 0:90 0:87 0:97 0:99
Ai D8 0:74 0:70 0:78 0:97 0:91 0:93 0:79
Ai D9 0:64 0:62 0:68 0:87 0:74 0:82 0:76
Ai D10 0:89 0:86 0:82 0:98 0:69 1:00 0:97
Ai D11 0:83 0:81 0:78 0:94 0:94 0:91 0:88
Ai D12 0:94 0:93 0:94 0:97 0:98 1:00 0:98
Ai D13 0:90 0:93 0:76 0:76 0:75 0:88 0:44
Ai D14 0:11 0:15 0:09 0:42 0:35 0:27 0:18
Ai D15 0:87 0:84 0:34 0:45 0:44 0:40 0:32
Ai D16 0:81 0:78 0:78 0:98 0:98 0:90 0:87
Ai D17 0:70 0:65 0:69 0:91 0:85 0:81 0:75
Average coeff. j1 .r/j 0:69 0:68 0:55 0.76 0.70 0.76 0:66
Overall coeff. j2 .r/j 0:14 0:31 0:18 0.68 0:44 0.71 0:43
Table 6.5 Pearson correlation coefficients under echo hiding

Individual Ai D1 0.96 0.96 0:96 1:00 1:00 0:99 1:00
coeff. Ai D2 0.99 0.99 0:99 1:00 0:98 1:00 0:99
.i; r/ Ai D3 0.97 0.97 0:99 1:00 0:97 0:99 1:00
Ai D4 0.99 0.99 0:99 1:00 0:99 1:00 0:99
Ai D5 0.98 0.97 0:99 1:00 1:00 0:99 1:00
Ai D6 0.93 0.98 0:91 0:85 0:84 0:90 0:83
Ai D7 0.94 0.94 0:99 1:00 0:99 0:99 1:00
Ai D8 0.98 0.98 0:99 0:99 1:00 0:99 0:99
Ai D9 0.96 0.96 0:95 1:00 0:98 0:98 1:00
Ai D10 0.94 0.93 0:96 0:99 0:96 0:98 0:99
Ai D11 0.99 0.99 1:00 0:98 0:98 1:00 0:98
Ai D12 0.95 0.96 0:98 1:00 0:99 0:99 1:00
Ai D13 0.02 0.16 0:13 0:19 0:14 0:17 0:12
Ai D14 0.71 0.71 0:72 0:79 0:80 0:77 0:79
Ai D15 0.97 0.97 0:97 1:00 1:00 0:99 1:00
Ai D16 0.97 0.97 0:99 1:00 0:99 0:99 1:00
Ai D17 0.97 0.96 0:99 1:00 1:00 0:99 1:00
Average coeff. j1 .r/j 0.90 0.90 0:91 0.93 0.92 0.92 0.92
Overall coeff. j2 .r/j 0.53 0.32 0:05 0.59 0:39 0.65 0:49
Table 6.6 Pearson correlation coefficients under histogram-based watermarking

Individual Ai D1 0:76 0:81 0:82 0:73 0:79 0:75 0:66
coeff. Ai D2 0:83 0:82 0:38 0:67 0:79 0:47 0:71
.i; r/ Ai D3 0:74 0:67 0:54 0:58 0:72 0:64 0:76
Ai D4 0:12 0:24 0:23 0:07 0:08 0:07 0:47
Ai D5 0:24 0:66 0:01 0:40 0:24 0:40 0:35
Ai D6 0:94 0:54 0:89 0:91 0:85 0:89 0:91
Ai D7 0:90 0:73 0:33 0:72 0:83 0:73 0:88
Ai D8 0:87 0:33 0:64 0:79 0:81 0:60 0:83
Ai D9 0:49 0:20 0:59 0:36 0:52 0:38 0:33
Ai D10 0:72 0:45 0:65 0:11 0:66 0:18 0:14
Ai D11 0:85 0:17 0:82 0:75 0:84 0:75 0:81
Ai D12 0:38 0:42 0:60 0:53 0:39 0:37 0:20
Ai D13 0:07 0:07 0:10 0:10 0:08 0:35 0:74
Ai D14 0:95 0:43 0:87 0:91 0:93 0:91 0:86
Ai D15 0:64 0:79 0:31 0:20 0:55 0:04 0:86
Ai D16 0:98 0:91 0:47 0:98 0:92 0:71 0:93
Ai D17 0:95 0:94 0:76 0:84 0:87 0:86 0:88
Average coeff. j1 .r/j 0.61 0:46 0:45 0:41 0.59 0:39 0:48
Overall coeff. j2 .r/j 0:40 0:43 0:28 0:35 0:22 0.71 0.50
over A1 , A2 , and A3 in the proposed audio watermarking scheme. It can be seen

that j2 .r/j, r D 1; 2; : : : ; 7 increase greatly to 0.84, 0.79, 0.85, 0.94, 0.81, 0.96,
and 0.94 respectively, not less than 0.79.
• Under different watermarking techniques, the LAR measure .r D 6/ shows the

best performance in both overall and average correlations. The LAR measure
provides the highest overall correlation under the proposed audio watermarking
scheme (j2 .6/j D 0:64), wavelet domain watermarking (j2 .6/j D 0:71), echo
hiding (j2 .6/j D 0:65), and histogram-based watermarking (j2 .6/j D 0:71).
Under cepstrum domain watermarking, the LAR measure yields the second
highest overall correlation (j2 .6/j D 0:70), only slightly less than the highest
value. Moreover, the LAR measure provides the highest average correlation
under the proposed scheme (j1 .6/j D 0:95) and wavelet domain watermarking
(j1 .6/j D 0:76). Also, the LAR measure yields the second highest average
correlation under cepstrum domain watermarking (j1 .6/j D 0:86) and echo
hiding (j1 .6/j D 0:92). However, the LAR measure receives the lowest average
correlation under histogram-based watermarking (j1 .6/j D 0:39).
After the LAR measure, the LLR measure .r D 4/ is also a good mea-
sure under different watermarking techniques. The LLR measure provides the
highest overall correlation under cepstrum domain watermarking (j2 .4/j D
0:72). Also, the LLR measure provides the second highest overall correla-
tion under wavelet domain watermarking (j2 .4/j D 0:68) and echo hiding
|ρ2(1)| = 0.84 |ρ2(2)| = 0.79 A1 : Soprano

0 0 A2 : Bass
−1 −1
A3 : Quartet
GQ
GQ
−2 −2
−3 −3
−4 −4
20 30 40 50 60 20 25 30 35
SNR measure segSNR measure
|ρ2(3)| = 0.85 |ρ2(4)| = 0.94

0 0
−1 −1
GQ
GQ
−2 −2
−3 −3
−4 −4
0 20 40 60 0 1 2 3 4
CD measure LLR measure
|ρ2(3)| = 0.81 |ρ2(4)| = 0.96

0 0
−1 −1
GQ
GQ
−2 −2
−3 −3
−4 −4
0 500 1000 0 5 10 15
LAR measure
IS measure
|ρ2(3)| = 0.94
0
−1
GQ
−2
−3
−4
0 5 10
WSS measure
Fig. 6.3 Overall correlation coefficients over audio test signals A1 , A2 , and A3
(j2 .4/j D 0:59). Moreover, the LLR measure provides the highest average
correlation under the proposed scheme (j1 .4/j D 0:95), cepstrum domain
watermarking (j1 .4/j D 0:87), wavelet domain watermarking (j1 .4/j D
0:76), and echo hiding (j1 .4/j D 0:93). However, the LLR measure receives
the second lowest average correlation under histogram-based watermarking
(j1 .4/j D 0:41).
In addition, the WSS measure .r D 7/ shows similar performance to the IS
measure .r D 5/, better than the SNR measure .r D 1/ and the segSNR measure
.r D 2/ on the whole.
By comparison, the CD measure .r D 3/ yields the worst correlation in most
cases, especially quite low overall correlation. The CD measure yields the lowest
overall correlation under the proposed scheme (j2 .3/j D 0:26), cepstrum
domain watermarking (j2 .3/j D 0:17), and echo hiding (j2 .3/j D 0:05).
6.4 Summary 175
Also, the CD measure yields the second lowest overall correlation under wavelet
domain watermarking (j2 .3/j D 0:18) and histogram-based watermarking
(j2 .3/j D 0:28).
• By using different quality measures, every audio watermarking technique can
achieve a satisfactory overall correlation. The highest overall correlation coeffi-
cient is equal to 0.64 under the proposed scheme, equal to 0.72 under cepstrum
domain watermarking, equal to 0.71 under wavelet domain watermarking, equal
to 0.65 under echo hiding, and equal to 0.71 under histogram-based watermark-
ing. As mentioned above, except for the fact that the highest overall correlation
under cepstrum domain watermarking is provided by the LLR measure, the
highest overall correlation under other watermarking techniques is provided by
the LAR measure.
This proves that objective quality measures are able to predict the perceptual
quality of the watermarked audio signals.
6.4 Summary
Imperceptibility is one prime concern in audio watermarking for copyrights pro-

tection. In this chapter, objective quality measures used in speech quality evaluation
have been assessed for their capabilities in the perceptual quality evaluation of audio
watermarking.
Different from perception modelling that mimics the human auditory system,
objective quality measures are adopted as an alternative approach to quantify
the dissimilarities caused by audio watermarking in an objective manner. Various
audio watermarking techniques discussed in Chaps. 3 and 4, such as our proposed
scheme, cepstrum domain watermarking, wavelet domain watermarking, echo
hiding, and histogram-based watermarking, are taken into consideration. During
the experiments using each technique, subjective listening tests and a commercial
evaluation tool PEMO-Q are used to grade the watermarked signals with different
watermark strengths. Moreover, the distances between the watermarked and host
signals are quantified by seven well-developed quality measures, i.e., the SNR,
segSNR, CD, LLR, IS, LAR, and WSS measures. Then two types of Pearson
correlation analyses are conducted to evaluate the performance of these quality
measures serving as the predictors for perceptual quality. For each quality measure,
one analysis is to calculate the average correlation coefficient by averaging the
corresponding individual correlation coefficients over different test signals, and the
other analysis is to calculate the overall correlation coefficient.
Pearson correlation coefficients show that the overall correlation coefficients
are commonly lower than the corresponding average correlation coefficients. It is
because the audio signals under test belong to different categories and possess differ-
ent time–frequency characteristics. Nevertheless, the investigated quality measures,
especially the LAR and the LLR measures, correlate well with the quasi-SDGs
from PEMO-Q. Moreover, quality measures run much faster than PEMO-Q. These
experimental results indicate that objective quality measures can be used reliably to
estimate the perceptual quality of the watermarked audio signals.
Appendix A
SDMI Standard
Secure digital music initiative (SDMI) is aiming at developing open technology

specifications that protect playing, storing, and distributing of digital music. To
achieve the goal of copyright enforcement, SDMI embeds robust and secure
watermarks (specifically called screening control data, SCD) into the music. Then
SDMI-compliant devices which are fitted with watermark detectors [145] can
identify the status of watermarks to perform appropriate operations.
In [146], key technical factors for evaluation were set forth in accordance with the
claimed performance of the technology, including inaudibility, robustness, reliabil-
ity, renewability, efficiency of operation, and effect on ability to compress content.
First, the requirement for inaudibility is that the content containing SCD should
be perceived as being statistically indistinguishable from the content prior to the
addition of SCD. For the robustness, the watermarked content must be able to
withstand each signal process listed in Table A.1. Here, the content is supposed to
be sampled at 44.1 or 48 kHz and quantized at 16 bits. Moreover, false-negative and
false-positive probabilities1 for reliability are required to be no less than 102 and
1012 , respectively. Third, renewability is expected to some extent, which indicates
that the technology should fail in limited ways after a successful attack and also
should provide a reasonable method of recovering from systematic compromise.
In addition, the technology must operate on a number of platforms to estimate the
efficiency of operation, i.e., measuring the amount of time necessary to detect or
embed SCD in relation to the length of the excerpt. Finally, the technology should
not interfere with the ability of standard compression algorithms to maintain an
expected fidelity level at standard bit rates.
1
False-negative probability is defined as the probability of missing detecting the existed watermark
and false-positive probability is defined as the probability of detecting the nonexisted watermark.
DOI 10.1007/978-3-319-07974-5, © Springer International Publishing Switzerland 2015
178 A SDMI Standard
Table A.1 Robustness test items in SDMI

Signal process Description
D/A, A/D D/A, A/D, converting twice
Equalization Typical case: 10-band graphic equalizer with the following
characteristics
Freq./Hz: 31 62 125 250 500 1 k 2 k 4 k 8 k
16 k
Gain/dB: 6 C6 6 C6 6 C6 6 C6
6 C6
Band-pass filtering 100–6 kHz, 12 dB/oct.
Linear speed change ˙10 %
Codecs (at typically ISO/IEC 13818-7: 1997 (“AAC”)
used data rates) ISO/IEC 14496-3: 1999 (MPEG-4 AAC with perceptual noise
substitution)
ISO/IEC 11172-3: 1993 Layer III (MPEG-1 Audio Layer 3
“MP3”)
Q-Design
Windows Media Audio
Twin-VQ
ATRAC-3
Dolby Digital AC-3 ATSC A_52
ePAC
Noise addition Adding white noise with constant level of 36 dB lower than
total averaged music power (S/N: 36 dB)
Time-scale modification Pitch-invariant time scaling: ˙4 %
Wow and flutter 0.5 % rms, from DC to 250 Hz
Addition echo Maximum delay: 100 ms
Feedback coefficient: up to 0.5
Down mixing and 6-channel to stereo
surround sound SRS
processing Spatializer
Dolby Surround
Dolby Headphone
Sample rate conversion 48 kHz ! 44.1 kHz
96 kHz ! 48/44.1 kHz
Dynamic range Threshold: 50 dB
reduction 16 dB max compression
Rate: 10 ms attack, 3 s recovery
Appendix B
STEP 2000
STEP 2000 [40] is a joint international evaluation project for audio digital water-
marking technology, undertaken by JASRAC1 and NRI2 together with international
associations of copyright management societies, CISAC and BIEM. It is the first
work of its kind initiated by copyright management bodies.
The objective of STEP 2000 is “to certify the aptitude of digital watermark
technologies, with a view towards promoting its utilization.” Enthusiastic responses
from many technology enterprises were received, contributing to an extensive
technology evaluation.
The evaluation of submitted digital watermark technologies was conducted
mainly with two aspects, i.e., audibility and robustness.
• Audibility—Whether the professionals can perceive if watermarks have been
embedded in music that is played back in a recording studio environment
Subjective listening test, ABX test, was conducted in perceptual quality evaluation.
First, the listener listens to a sound recording with no watermark (A), a sound
recording with watermarks embedded (B), and a sound recording which is one of
the two (X). After that, a listener listens to A and B alternately twice for 40 s each
and listens to X for 40 s again. Then, the listener decides whether X is A or B.
There are two requirements in ABX test to ensure its validity. One is to eliminate
(correct) contingency responses. To this end, the above tests were conducted
five times for each system. Moreover, the listener is defined to have detected
the embedded watermark if the same listener correctly determines whether the
watermark is embedded or not on each of the five tests. Under this definition,
significance of the responses are 95 % or greater. The other is to ensure typicality of
the professionals from the recording industry. For this purpose, a group comprising
of one recording engineer, one mastering engineer, one synthesizer manipulator, and
one audio critic was selected.
1
JASRAC: Japanese Society for Rights of Authors, Composers and Publishers
2
NRI: Nomura Research Institute, Ltd.
180 B STEP 2000
Table B.1 Robustness test items in STEP 2000

Testing item Overview of processing involved
D/A, A/D transition Digital!Analog!Digital
Altered number of channels Stereo (2ch)!mono
Down sampling 44.1 kHz/16 bit/2ch!16 kHz/16 bit/2ch
Amplitude compression 44.1 kHz/16 bit/2ch!44.1 kHz/8 bit/2ch
Time and pitch compression Time compression/decompression: ˙10 %
and decompression Pitch shift compression/decompression: ˙10 %
Linear data compression MPEG 1 Audio Layer 3 (MP3): 128 kbps
MPEG 2 AAC: 128 kbps
ATRAC: Version 4.5
ATRAC 3: 105 kbps
RealAudio: ISDN
Windows Media Audio: ISDN
Nonlinear data compression FM (FM multiple broadcast, terrestrial hertzian TV broadcast)
AM (AM broadcast)
PCM (Satellite TV broadcast: communications satellite,
broadcasting satellite)
Characteristic transformation FM (FM multiple broadcast, terrestrial hertzian TV broadcast)
of frequency response AM (AM broadcast)
PCM (Satellite TV broadcast: communications satellite,
broadcasting satellite)
Noise White noise: S=N D 40 dB
• Robustness—Whether the watermarked data can be extracted after various

processes of music usage
The robustness tests were performed under equal conditions for all the submitted
technologies. In general, watermarked data were manipulated in some way, for
example, processed in a mastering studio, processed in a broadcasting studio (and
a prospective broadcasting environment), processed for distribution through the
Internet and other networks, and processed by commonly available consumer level
equipments. Say specifically, robustness testing items are listed in Table B.1.
Appendix C
StirMark for Audio
StirMark for Audio [134] is a generic tool of robustness test for audio watermarking
systems. It is derived from StirMark,1 a fair benchmark for image watermarking. A
number of attacks as well as attack parameters are included in StirMark for Audio
v0.2, as shown in Table C.1.
Table C.1 Robustness test items in StirMark for Audio

Attack name Description Parameter used
AddBrumm Add buzz or sinus tone to the sound. The unit of AddBrummfrom
the three values is samples and for the AddBrummto
frequency hertz (Hz) AddBrummstep
AddBrummFreq
AddDynNoise Add a dynamic white noise part to the samples. Dynnoise
The given parameter sets the maximum
noise value
AddFFTNoise Add white noise to the samples in the FFT FFTSIZE
domain. The value “FFTNoise” sets the FFTNoise
power of this attack to add the noise
AddNoise Add white noise to the samples. The unity is in Noisefrom
sample values. The value “0” adds nothing Noiseto
and “32768” the absolute distorted Noisestep
maximum
(continued)
1
StirMark v3.1 is a first benchmark for image watermarking released in 1999. The latest version is
StirMark Benchmark 4.0, available at http://www.petitcolas.net/fabien/watermarking/stirmark/.
182 C StirMark for Audio
Table C.1 (continued)

AddSinus Add a sinus signal to the sound file. With this AddSinusFreq
attack you can insert a disturb signal in the AddSinusAmp
frequency band where the watermark is
located. The unit of the frequency parameter
is hertz (Hz) and samples
Amplify Change the loudness of audio file. For example Amplify
the value “100” does not change the
amplitude and a value “50” means a half
loudness
Compressor This attack works like a compressor. You can ThresholdDB
increase or decrease the loudness of CompressValue
passages. The unit of the threshold is decibel
(dB). The “CompressValue” describes how
the sample can be changed. “2” means that
the loudness of all samples in the threshold
will be half. If the value is less than “1,” the
compressor is an expander and will increase
the loudness
CopySample Similar to the FlippSample attack, but this Period
attack copies the samples between the FlippDist
samples with a distance of FlippDist FlippCount
CutSamples Remove samples from the audio file. If the value Remove
of “Remove” is “10000,” then this attack RemoveNumber
removes every “10000” samples
“RemoveNumber” samples periodically
Echo Add an echo to the sound file. The given value Period
means the distance of the echo
Exchange Swap two sequent samples for all samples
ExtraStereo Increase the stereo part of the file. If the file ExtraStereofrom
does not have a stereo part (only mono), then ExtraStereoto
this attack does not have an effect ExtraStereostep
FFT_HLPassQuick Similar to the RC-HighPass and RC-LowPass FFTSIZE
attacks, but this attack is performed in the HighPassFreq
FFT domain. FFT window size can be set LowPassFreq
with the “FFTSIZE” parameter. This attack
does not fade between the FFT windows, so
it is possible to hear knocks
FFT_Invert Invert all samples (real and imaginary part) in FFTSIZE
the FFT domain
C StirMark for Audio 183
Table C.1 (continued)

FFT_RealReverse Reverse only the real part from the FFT FFTSIZE
FFT_Stat1 Statistical attack in the FFT domain
FFT_Test Currently, it is to swap some samples inside FFTSIZE
from FFT
FlippSample Swap samples inside the sound file periodically. Period
It swaps every “Period” “FlippCount” FlippCount
samples with samples which have a distance FlippDist
of “FlippDist”
Invert Invert all samples in the audio file
LSBZero This attack sets all LSB to “0” (zero)
Normalize Normalize the amplitude to the maximum value
Nothing This attack does nothing with the audio file. The
watermark should be retrieved perfectly
RC-HighPass Simulate a high-pass filter built with a resistance HighPassFreq
(R) and a capacitor (C)
RC-LowPass Simulate a low-pass filter like RC-HighPass LowPassFreq
Resampling Change the sampling rate of sound file SampleRate
Smooth This attack smoothes the samples. The setting
sample value depends on the samples before
and after the modifying point
Smooth2 Similar to Smooth, but the neighbor samples are
valued a little bit different
Stat1
Stat2
VoiceRemove Is the opposite to ExtraStereo. This attack
removes the mono part of the file (mostly
where the voice is). If the file does not have
a stereo part (only mono), then everything
will be removed
ZeroCross This attack likes a limiter. If the sample value is ZeroCross
less than the given value (threshold), all
samples are set to zero
ZeroLength If a sample value is exactly “0” (zero), then this ZeroLength
attack inserts more samples with the value
“0” (zero)
ZeroRemove This attack removes all samples where the value
is “0” (zero)
Appendix D
Critical Bandwidth
See Table D.1.
Table D.1 Critical bands over the frequency spectrum [11]

Critical band rate Lower frequency Upper frequency Center frequency Critical bandwidth
z=Bark fl =Hz fh =Hz fc =Hz f =Hz
0 0 100 50 100
1 100 200 150 100
2 200 300 250 100
3 300 400 350 100
4 400 510 450 110
5 510 630 570 120
6 630 770 700 140
7 770 920 840 150
8 920 1,080 1,000 160
9 1,080 1,270 1,170 190
10 1,270 1,480 1,370 210
11 1,480 1,720 1,600 240
12 1,720 2,000 1,850 280
13 2,000 2,320 2,150 320
14 2,320 2,700 2,500 380
15 2,700 3,150 2,900 450
16 3,150 3,700 3,400 550
17 3,700 4,400 4,000 700
18 4,400 5,300 4,800 900
19 5,300 6,400 5,800 1,100
20 6,400 7,700 7,000 1,300
21 7,700 9,500 8,500 1,800
22 9,500 12,000 10,500 2,500
23 12,000 15,500 13,500 3,500
24 15,500
Appendix E
List of Audio Test Files
All the audio test samples are 44.1 kHz, 16 bit, monaural wave format files, as listed
in Table E.1.
Table E.1 Descriptions of audio test files for performance evaluation
188 E List of Audio Test Files
Appendix F
Basic Robustness Test
Basic robustness test is applied to the watermarked audio signal Sw to inspect its
capability of resisting different attacks as listed in Table F.1.
Table F.1 Descriptions of basic robustness test

Parameters (default
Testing item value) Expression Implementation
No attack / No attack /
Noise addition snr: signal-to-noise Noise (snr) In MATLAB
ratio (36 dB)
Resampling fw : downsampling Resampling (fw ) Adobe Audition v3.0:
frequency 44:1 kHz!fw !44:1 kHz
(22.05 kHz)
Requantization Qw : requantization bit Requantization (Qw ) Adobe Audition v3.0:
number (8 bit) 16 bit ! Qw ! 16 bit
Amplitude scaling As : rate of scaling Amplitude (˙As ) Adobe Audition v3.0
(10 %)
Lowpass filtering fcutoff : cutoff Lp filtering (fcutoff ) In MATLAB
frequency
(8 kHz)
DA/AD R: recording mode DA/AD (R) Adobe Audition v3.0:
conversion (line-in jack) play and record
Echo addition Am : normalized Echo (Am ; td ) In MATLAB
amplitude
attenuation (0.3)
td : delay time
(200 ms)
Reverberation treverb : reverberation Reverb (treverb ) Adobe Audition v3.0
time (1 s)
MP3 compression m: compression Compression I (m) Adobe Audition v3.0:
bitrate Compression :wav ! :mp3 ! :wav
(96 kbps) II (m)
(continued)
190 F Basic Robustness Test
Table F.1 (continued)

Parameters (default
Testing item value) Expression Implementation
Random samples nc : no. of croppings (8) Cropping (nc tc ) In MATLAB
cropping tc : cropped interval
(25 ms)
Jittering tj : cut clips (0.1 ms) Jittering (tj =tf ) In MATLAB
tf : interval of cutting
(20 ms)
Zeros inserting nz : no. of insertings (8) Inserting (nz tz ) In MATLAB
tz : inserted interval
(25 ms)
Pitch-invariant PTSM : percentage of PITSM (PTSM ) Adobe Audition v3.0
time-scale time stretching (˙4 %)
modification
Tempo-preserved PPSM : percentage of TPPSM (PPSM ) Adobe Audition v3.0
pitch-scale pitch shifting (˙4 %)
modification
Appendix G
Nonuniform Subbands
Audio signals used in this book are in WAVE format (44.1 kHz, 16 bit), and hence
the nonuniform subbands are designed to cover 100 22; 050 Hz frequency band.
When Nsubband D 32, the lower/upper limits of the subbands obtained are presented
in Table G.1. The number of FFT coefficients in each subband is calculated based
on a frame length N D 512.
Table G.1 Thirty-two nonuniform subbands over the frequency spectrum

Subband Lower limit Upper limit Bandwidth No. of FFT
index V l /Hz V h /Hz Bw /Hz coefficients
1 559.8 990.5 430.7 5
2 990.5 1,421.2 430.7 5
3 1,421.2 1,851.9 430.7 5
4 1,851.9 2,282.6 430.7 5
5 2,282.6 2,713.3 430.7 5
6 2,713.3 3,144.0 430.7 5
7 3,144.0 3,574.7 430.7 5
8 3,574.7 4,005.4 430.7 5
9 4,005.4 4,436.1 430.7 5
10 4,436.1 4,866.8 430.7 5
11 4,866.8 5,297.5 430.7 5
12 5,297.5 5,728.2 430.7 5
13 5,728.2 6,158.9 430.7 5
14 6,158.9 6,589.6 430.7 5
15 6,589.6 7,020.3 430.7 5
16 7,020.3 7,537.1 516.8 6
17 7,537.1 8,053.9 516.8 6
18 8,053.9 8,570.7 516.8 6
19 8,570.7 9,173.6 602.9 7
20 9,173.6 9,776.5 602.9 7
21 9,776.5 10,465.6 689.1 8
22 10,465.6 11,240.8 775.2 9
(continued)
192 G Nonuniform Subbands
Table G.1 (continued)

Subband Lower limit Upper limit Bandwidth No. of FFT
index V l /Hz V h /Hz Bw /Hz coefficients
23 11,240.8 12,016.0 775.2 9
24 12,016.0 12,791.2 775.2 9
25 12,791.2 13,738.7 947.5 11
26 13,738.7 14,686.2 947.5 11
27 14,686.2 15,633.7 947.5 11
28 15,633.7 16,753.4 1,119.7 13
29 16,753.4 17,873.1 1,119.7 13
30 17,873.1 19,165.1 1,292.0 15
31 19,165.1 20,457.1 1,292.0 15
32 20,457.1 21,835.2 1,378.1 16
References
1. H. Malik, R. Ansari, A. Khokhar, Robust audio watermarking using frequency-selective

spread spectrum. IET Inform. Secur. 2(4), 129–150 (2008)
2. S.J. Xiang, H.J. Kim, J.W. Huang, Audio watermarking robust against time-scale modification
and mp3 compression. Signal Process. 88(10), 2372–2387 (2008)
3. O.T.-C. Chen, W.-C. Wu, Highly robust, secure, and perceptual-quality echo hiding scheme.
IEEE Trans. Audio Speech Lang. Process. 16(3), 629–638 (2008)
4. X. He, Watermarking in Audio: Key Techniques and Technologies (Cambria Press,
Youngstown, 2008)
5. W. Li, X. Y. Xue, P.Z. Lu, Localized audio watermarking technique robust against time-scale
modification. IEEE Trans. Multimed. 8(1), 60–69 (2006)
6. M.F. Mansour, A.H. Tewfik, Data embedding in audio using time-scale modification. IEEE
Trans. Speech Audio Process. 13(3), 432–440 (2005)
7. N. Cvejic, T. Seppanen, Robust audio watermarking in wavelet domain using frequency
hopping and patchwork method, in Proceedings of the 3rd International Symposium on Image
and Signal Processing and Analysis, 2003, pp. 251–255
8. D. Kirovsk, H.S. Malvar, Spread-spectrum watermarking of audio signals. IEEE Trans. Signal
Process. 51(4), 1020–1033 (2003)
9. R. Tachibana, S. Shimizu, S. Kobayashi, An audio watermarking method using a two-
dimensional pseudo-random array. Signal Process. 82(10), 1455–1469 (2002)
10. R. Tachibana, Improving audio watermarking robustness using stretched patterns against
geometric distortion, in Proceedings of IEEE Pacific-Rim Conference on Multimedia (PCM),
2002, pp. 647–654
11. E. Zwicker, H. Fastl, Psychoacoustics: Facts and Models (Springer, Berlin, 1990)
12. G. Widmer, D. Rocchesso, V. Välimäki, C. Erkut, F. Gouyon, D. Pressnitzer, et al., Sound
and music computing: research trends and some key issues. J New Music Res. 36, 169–184
(2007)
13. I.J. Cox, M.L. Miller, J.A. Bloom, J. Fridrich, T. Kalker, Digital Watermarking and
Steganography (Morgan Kaufmann Publishers, San Francisco, 2008)
14. S. Katzenbeisser, F.A.P. Petitcolas (eds.), Information Hiding Techniques for Steganography
and Digital Watermarking (Artech House, Boston, 2000)
15. N.F. Johnson, Z. Duric, S. Jajodia, Information Hiding: Steganography and Watermarking -
Attacks and Countermeasures (Kluwer Academic, Boston, 2001)
16. R. Walker, Audio watermarking. Technical Report, BBC R&D (2004) [Online], http://www.
bbc.co.uk/rd/pubs/whp/whp-pdf-files/WHP057.pdf
17. S.P. Mohanty, Digital watermarking: a tutorial review. Technical Report, University
of South Florida (1999) [Online], http://www.cs.unt.edu/smoh-anty/research/Reports/
MohantyWatermarkingSurvey1999.pdf
194 References
18. M.D. Swanson, M. Kobayashi, A.H. Tewfik, Multimedia data: embedding and watermarking
technologies. Proc. IEEE 86(6), 1064–1087 (1998)
19. Y.Q. Lin, W.H. Abdulla, Audio watermarking for copyrights protection. Technical Report
SoE-650, School of Engineering, The University of Auckland (2007)
20. F.A.P. Petitcolas, R.J. Anderson, M.G. Kuhn, Information hiding: a survey. Proc. IEEE 87(7),
1062–1078 (1999)
21. S.A. Craver, M. Wu, B. Liu, What can we reasonably expect from watermarks? in Proceedings
of IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, 2001, pp.
223–226
22. L.d.C.T. Gomes, P. Cano, E. Gómez, M. Bonnet, E. Batlle, Audio watermarking and
fingerprinting: for which applications? J. New Music Res. 32(1), 65–81 (2003)
23. F. Kurth, M. Muller, Efficient index-based audio matching. IEEE Trans. Audio Speech Lang.
Process. 16(2), 382–395 (2008)
24. M. Barni, F. Bartolini, Watermarking Systems Engineering: Enabling Digital Assets Security
and Other Applications (Marcel Dekker, New York, 2004)
25. J.-S. Pan, H.-C. Huang, L.C. Jain (eds.), Intelligent Watermarking Techniques (World
Scientific, River Edge, 2004)
26. J. Seitz (ed.), Digital Watermarking for Digital Media (Information Science Publishers,
Hershey, 2005)
27. B. Furht, D. Kirovski (eds.), Multimedia Watermarking Techniques and Applications (Auer-
bach Publications, Boca Raton, 2006)
28. F. Hartung, M. Kutter, Multimedia watermarking techniques. Proc. IEEE 87(7), 1079–1107
(1999)
29. T. Page, Digital watermarking as a form of copyright protection. Comput. Law Secur. Rep.
14(6), 390–392 (1998)
30. N. Cvejic, T. Seppanen (eds.), Digital Audio Watermarking Techniques and Technologies:
Applications and Benchmarks (Information Science Reference, Hershey, 2008)
31. M. Arnold, M. Schmucker, S.D. Wolthusen, Techniques and Applications of Digital Water-
marking and Content Protection (Artech House, Boston, 2003)
32. L. Boney, A.H. Tewfik, K.N. Hamdy, Digital watermarks for audio signals, in Proceedings of
IEEE International Conference on Multimedia Computing and Systems, 1996, pp. 473–480
33. C.-P. Wu, P.-C. Su, C.-C.J. Kuo, Robust and efficient digital audio watermarking using audio
content analysis, in Proceedings of SPIE Security and Watermarking of Multimedia Contents
II, vol. 3971, 2000 , pp. 382–392
34. W.-N. Lie, L.-C. Chang, Robust and high-quality time-domain audio watermarking subject to
psychoacoustic masking, in Proceedings of IEEE International Symposium on Circuits and
Systems (ISCAS), 2001, pp. 45–48
35. S.J. Xiang, J.W. Huang, Histogram-based audio watermarking against time-scale modification
and cropping attacks. IEEE Trans. Multimed. 9(7), 1357–1372 (2007)
36. W. Bender, D. Gruhl, N. Morimoto, A. Lu, Techniques for data hiding. IBM Syst. J. 35(3 &
4), 313–336 (1996)
37. F.A.P. Petitcolas, Watermarking schemes evaluation. IEEE Signal Process. Mag. 17(5), 58–64
(2000)
38. A. Lang, J. Dittmann, Transparency and complexity benchmarking of audio watermark-
ing algorithms issues, in Proceedings of Workshop on Multimedia and Security, 2006,
pp. 190–201
39. M. Arnold, Subjective and objective quality evaluation of watermarked audio tracks, in
Proceedings of International Conference on Web Delivering of Music (WEDELMUSIC),
2002, pp. 161–167
40. Announcement of Evaluation Test Results for “STEP 2000". JASRAC and NRI (2000)
[Online], http://www.jasrac.or.jp/watermark/ehoukoku.htm
41. A. Garay Acevedo, Audio watermarking quality evaluation, in e-Business and Telecommuni-
cation Networks, ed. by J. Ascenso et al. (Springer, Netherlands, 2006), pp. 272–283
References 195
42. G. Stoll, F. Kozamernik, EBU listening tests on internet audio codecs. EBU Technical Review,
2000
43. ITU-R Recommendation BS.1116-1, ITU-R Recommendation BS.1116-1: Methods for the
Subjective Assessment of Small Impairments in Audio Systems Including Multichannel Sound
Systems, 1997
44. ITU-R Recommendation BS.1534-1, ITU-R Recommendation BS.1534-1: Method for the
Subjective Assessment of Intermediate Quality Level of Coding Systems, 2003
45. ITU-R Recommendation BS.1284-1, ITU-R Recommendation BS.1284-1: General methods
for the subjective assessment of sound quality, 2003
46. J.G. Beerends, Audio quality determination based on perceptual measurement techniques,
in Applications of Digital Signal Processing to Audio and Acoustics, ed. by M. Kahrs,
K. Brandenburg (Kluwer Academic, Boston, 1998), pp. 1–38
47. A. Lerch, Software: EAQUAL - Evaluation of Audio Quality, v.0.1.3alpha ed. (2002)
[Online], http://www.rarewares.org/others.php
48. P. Kabal, An examination and interpretation of ITU-R BS.1387: Perceptual evaluation of
audio quality. Technical Report, TSP Lab, McGill University (2003) [Online], http://www-
mmsp.ece.mcgill.ca/Documents
49. R. Huber, B. Kollmeier, PEMO-Q: a new method for objective audio quality assessment
using a model of auditory perception. IEEE Trans. Audio Speech Lang. Process. 14(6),
1902–1911 (2006) [Online]. http://www.hoertech.de/web-en/produkte/downloads.shtml
50. S.R. Quackenbush, T.P. Barnwell III, M.A. Clements, Objective Measures of Speech Quality
(Prentice Hall, Englewood Cliffs, 1988)
51. M. Bosi, R.E. Goldberg, Introduction to Digital Audio Coding and Standards (Kluwer
Academic, Boston, 2003)
52. W.J. Vincoli (ed.), Lewis’ Dictionary of Occupational and Environmental Safety and Health
(Lewis Publishers, Boca Raton, 2000)
53. K. Johnson, Acoustic and Auditory Phonetics (Blackwell Publisher, Malden, 2003)
54. P.H. Lindsay, D.A. Norman, Human Information Processing: An Introduction to Psychology
(Academic, New York, 1977)
55. T.S. Gunawan, Audio compression and speech enhancement using temporal masking models.
Ph.D. dissertation, The University of New South Wales, 2007
56. [Online]. Available: http://projects.cbe.ab.ca/Diefenbaker/Biology/Bio%20Website
%20Final/notes/nervous_system/Image59.gif
57. M.W. Levine, Levine and Shefner’s Fundamentals of Sensation and Perception (Oxford
University Press, Oxford, 2000)
58. W.A. Yost, D.W. Nielsen, Fundamentals of Hearing: An Introduction (Holt, Rinehart and
Winston, New York, 1977)
59. E.A.G. Shaw, Earcanal pressure generated by a free sound field. J. Acoust. Soc. Am. 39(3),
465–470 (1966)
60. B.C.J. Moore, An Introduction to the Psychology of Hearing (Academic, New York, 2003)
61. [Online]. Available: http://www.chicagoear.com/images/earworks.gif
62. T.D. Rossing (ed.), Handbook of Acoustics (Springer, Heidelberg, 2007)
63. [Online]. Available: http://www2.ph.ed.ac.uk/AardvarkDeployments/Public/67158/
views/workspace/dwatts1/66265/inner.node/les/MusicalAcoustics/CourseNotes/
PropertiesoftheEar/web.html
64. [Online]. Available: http://www.ai.rug.nl/acg/cpsp/docs/cochleaModel.html
65. I.J. Hirsh, The Measurement of Hearing (McGraw-Hill, New York, 1952)
66. H. Fletcher, W.A. Munson, Loudness, its definition, measurement and calculation. J. Acoust.
Soc. Am. 5(2), 82–108 (1933)
67. Y.H. Kim, H.I. Kang, K.I. Kim, S.-S. Han, A digital audio watermarking using two masking
effects, in Advances in Multimedia Information Processing - PCM 2002, ed. by Y.-C.
Chen, L.-W. Chang, H.C.-T. Lecture Notes in Computer Science, vol. 2532 (Springer,
Berlin/Heidelberg, 2002), pp. 105–115
196 References
68. X.M. Quan, H.B. Zhang, Statistical audio watermarking algorithm based on perceptual
analysis, in Proceedings of the 5th ACM Workshop on Digital Rights Management, 2005,
pp. 112–118
69. E. Ambikairajah, A.G. Davis, W.T.K. Wong, Auditory masking and MPEG-1 audio compres-
sion. Electron. Comm. Eng. J. 9, 165–173 (1997)
70. A. Spanias, T. Painter, V. Atti, Audio Signal Processing and Coding (Wiley-Interscience,
Hoboken, 2007)
71. M.D. Swanson, B. Zhu, A.H. Tewfik, L. Boney, Robust audio watermarking using perceptual
masking. Signal Process. 66(3), 337–355 (1998)
72. R.A. Garcia, Digital watermarking of audio signals using a psychoacoustic auditory model
and spread spectrum theory. AES E-Library, 1999
73. S. Ratanasanya, S. Poomdaeng, S. Tachphetpiboon, T. Amornraksa, New psychoacoustic
models for wavelet based audio watermarking, in IEEE International Symposium on Com-
munications and Information Technology (ISCIT), vol. 1, pp. 602–605, 2005
74. ISO/IEC IS 11172-3, Information Technology - Coding of Moving Picture and Associated
Audio for Digital Storage Media Up To About 1.5Mbit/s, Part 3: Audio (BSI, London, 1993)
75. K.C. Pohlmann, Principles of Digital Audio (McGraw-Hill, New York, 2000)
76. D. Pan, A tutorial on MPEG/audio compression. IEEE Multimed. 2, 60–74 (1995)
77. F.A.P. Petitcolas, MPEG for Matlab, v.1.2.8 ed. (2003) [Online], http://www.petitcolas.net/
fabien/software/mpeg
78. C.-Y. Lin, An investigation into perceptual audio coding and the use of auditory gammatone
filterbanks. Master’s thesis, The University of Auckland, 2007
79. SQAM - Sound Quality Assessment Material, European Broadcasting Union (EBU)
[Online], http://sound.media.mit.edu/mpeg4/audio/sqam
80. A. Takahashi, R. Nishimura, Y. Suzuki, Multiple watermarks for stereo audio signals using
phase-modulation techniques. IEEE Trans. Signal Process. 53(2), 806–815 (2005)
81. P. Liew, M. Armand, Inaudible watermarking via phase manipulation of random frequencies.
Multimed. Tools Appl. 35(3), 357–377 (2007)
82. A. Piva, M. Barni, F. Bartolini, A. De Rosa, Data hiding technologies for digital radiography.
IEE Proc. Vision Image Signal Process. 152(5), 604–610 (2005)
83. B. Chen, G.W. Wornell, Quantization index modulation: a class of provably good methods
for digital watermarking and information embedding. IEEE Trans. Inform. Theory 47(4),
1423–1443 (2001)
84. A. Zaidi, R. Boyer, P. Duhamel, Audio watermarking under desynchronization and additive
noise attacks. IEEE Trans. Signal Process. 54(2), 570–584 (2006)
85. D. Lam, Audio watermarking. COMPSYS401A Project, The University of Auckland, 2003
86. S. Saito, T. Furukawa, K. Konishi, A digital watermarking for audio data using band division
based on QMF bank, in Proceedings of IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), vol. 4, 2002, pp. 3473–3476
87. A.V. Oppenheim, R.W. Schafer, Discrete-Time Signal Processing (Prentice Hall, Englewood
Cliffs, 1989)
88. S.-S. Kuo, J.D. Johnston, W. Turin, S.R. Quackenbush, Covert audio watermarking using per-
ceptually tuned signal independent multiband phase modulation. Proc. ICASSP 2, 1753–1756
(2002)
89. I.J. Cox, J. Kilian, F.T. Leighton, T. Shamoon, Secure spread spectrum watermarking for
multimedia. IEEE Trans. Image Process. 6(12), 1673–1687 (1997)
90. H.J. Kim, Audio watermarking techniques, in Proceedings of Pacific Rim Workshop on Digital
Steganography, 2003
91. H. Malik, A. Khokhar, A. Rashid, Robust audio watermarking using frequency selective
spread spectrum theory. Proc. ICASSP 5, 385–388 (2004)
92. N. Cvejic, T. Seppanen, Spread spectrum audio watermarking using frequency hopping and
attack characterization. Signal Process. 84(1), 207–213 (2004)
93. J. Seok, J. Hong, J. Kim, A novel audio watermarking algorithm for copyright protection of
digital audio. ETRI J. 24(3), 181–189 (2002)
References 197
94. L.R. Rabiner, R.W. Schafer, Digital Processing of Speech Signals (Prentice Hall, Englewood
Cliffs, 1978)
95. X. Li, H.H. Yu, Transparent and robust audio data hiding in cepstrum domain, in Proceedings
of IEEE International Conference on Multimedia and Expo (ICME), vol. 1, 2000, pp. 397–400
96. S.-K. Lee, Y.-S. Ho, Digital audio watermarking in the cepstrum domain. IEEE Trans.
Consumer Electron. 46(3), 744–750 (2000)
97. C.-T. Hsieh, P.-Y. Sou, Blind cepstrum domain audio watermarking based on time energy
features, in Proceedings of International Conference on Digital Signal Processing (DSP),
vol. 2, 2002, pp. 705–708
98. L.L. Cui, S.X. Wang, T.F. Sun, The application of binary image in digital audio watermarking,
in Proceedings of International Conference on Neural Networks and Signal Processing,
vol. 2, 2003, pp. 1497–1500
99. K. Gopalan, Audio steganography by cepstrum modification. Proc. ICASSP 5, 481–484
(2005)
100. K. K. Parhi, T. Nishitani, Digital Signal Processing for Multimedia Systems (CRC Press, New
York, 1999)
101. W.Y. Hwang, H.I. Kang, S.S. Han, K.I. Kim, H.S. Kang, Robust audio watermarking using
both DWT and masking effect, in Digital Watermarking, LNCS 2939, ed. by T. Kalker et al.
(Springer, Berlin/Heidelberg, 2004), pp. 382–389
102. A. Prochazka, J. Uhlir, P.W.J. Rayner, N.G. Kingsbury, Signal Analysis and Prediction
(Birkhäuser, Boston, 1998)
103. X. He, M.S. Scordilis, An enhanced psychoacoustic model based on the discrete wavelet
packet transform. J. Franklin Inst. 343(7), 738–755 (2006)
104. C.-S. Ko, K.-Y. Kim, R.-W. Hwang, Y.-S. Kim, S.-B. Rhee, Robust audio watermarking
in wavelet domain using pseudorandom sequences, in Proceedings of Annual International
Conference on Computer and Information Science (ACIS), 2005, pp. 397–401
105. P. Artameeyanant, Wavelet audio watermark robust against MPEG compression, in SICE
Annual Conference, pp. 1414–1417, 2007
106. H.O. Kim, B.K. Lee, N. Lee, Wavelet-based audio watermarking techniques: robustness and
fast synchronization [Online], http://amath.kaist.ac.kr/research/paper/01-11.pdf
107. W. Li, X.Y. Xue, An audio watermarking technique that is robust against random cropping.
Comput. Music J. 27(4), 58–68 (2003)
108. H.O. Oh, J.W. Seok, J.W. Hong, D.H. Youn, New echo embedding technique for robust and
imperceptible audio watermarking. Proc. ICASSP 3, 1341–1344 (2001)
109. D. Gruhl, A. Lu, W. Bender, Echo hiding, in Information Hiding, ed.b y R. Anderson. Lecture
Notes in Computer Science, vol. 1174 (Springer, Berlin/Heidelberg, 1996), pp. 295–315
110. H.J. Kim,Y.H. Choi, A novel echo-hiding scheme with backward and forward kernels. IEEE
Trans. Circ. Syst. Video Tech. 13(8), 885–889 (2003)
111. B.-S. Ko, R. Nishimura, Y. Suzuki, Time-spread echo method for digital audio watermarking.
IEEE Trans. Multimed. 7(2), 212–221 (2005)
112. B.-S. Ko, R. Nishimura, Y. Suzuki, Log-scaling watermark detection in digital audio
watermarking. Proc. ICASSP 3, 81–84 (2004)
113. D. Coltuc, P. Bolon, Robust watermarking by histogram specification, in Proceedings of
International Conference on Image Processing (ICIP), vol. 2, 1999, pp. 236–239
114. M. Mese, P.P. Vaidyanathan, Optimal histogram modification with MSE metric. Proc.
ICASSP 3, 1665–1668 (2001)
115. E. Chrysochos, V. Fotopoulos, A.N. Skodras, M. Xenos, Reversible image watermarking
based on histogram modification, in Proceedings of the 11th Panhellenic Conference on
Informatics (PCI), vol. B, 2007, pp. 93–104
116. G.R. Xuan, Q.M. Yao, C.Y. Yang, J.J. Gao, P.Q. Chai, Y. Shi, Z.C. Ni, Lossless data
hiding using histogram shifting method based on integer wavelets, in Digital Watermark-
ing, ed. by Y.Q. Shi, B. Jeon. Lecture Notes in Computer Science, vol. 4283 (Springer,
Berlin/Heidelberg, 2006), pp. 323–332
198 References
117. S.J. Xiang, J.W. Huang, R. Yang, Time-scale invariant audio watermarking based on the
statistical features in time domain, in Information Hiding, ed. by J. Camenisch et al. Lecture
Notes in Computer Science, vol. 4437 (Springer, Berlin/Heidelberg, , 2007), pp. 93–108.
Matlab implementation available at http://cist.korea.ac.kr/xiangshijun/
118. D.R. Smith, Digital Transmission Systems (Kluwer Academic, Boston, 2004)
119. H. Farid, Detecting hidden messages using higher-order statistical models. Proc. ICIP 2,
905–908 (2002)
120. M. Alghoniemy, A.H. Tewfik, Image watermarking by moment invariants. Proc. ICIP 2,
73–76 (2000)
121. S.J. Xiang, J.W. Huang, R. Yang, C.T. Wang, H.M. Liu, Robust audio watermarking based on
low-order zernike moments, in Digital Watermarking, ed. by Y.Q. Shi, B. Jeon. Lecture Notes
in Computer Science, vol. 4283 (Springer, Berlin/Heidelberg, 2006), pp. 226–240
122. P. Bas, J.-M. Chassery, B. Macq, Geometrically invariant watermarking using feature points.
IEEE Trans. Image Process. 11(9), 1014–1028 (2002)
123. F.-S. Wei, F. Xue, M.Y. Li, A blind audio watermarking scheme using peak point extraction.
Proc. ISCAS 5, 4409–4412 (2005)
124. W.H. Abdulla, Auditory based feature vectors for speech recognition systems, in Advances
in Communications and Software Technologies, ed. by N.E. Mastorakis, V.V. Kluev (WSEAS
Press, Greece, 2002), pp. 231–236
125. Y.Q. Lin, W.H. Abdulla, Robust audio watermarking technique based on Gammatone
filterbank and coded-image, in Proceedings of International Symposium on Signal Processing
and Its Applications (ISSPA), 2007
126. D. Bailey, W. Cammack, J. Guajardo, C. Paar, Cryptography in modern communication
systems, in TI DSPS FEST, pp. 1–15, 1999
127. Y.Q. Lin, W.H. Abdulla, A secure and robust audio watermarking scheme using multiple
scrambling and adaptive synchronization, in Proceedings of the 6th International Conference
on Information, Communications and Signal Processing (ICICS), 2007
128. Y.Q. Lin, W.H. Abdulla, Y. Ma, Audio watermarking detection resistant to time and pitch
scale modification, in Proceedings of IEEE International Conference on Signal Processing
and Communications (ICSPC), 2007, pp. 1379–1382
129. M. Kahrs, K. Brandenburg, Applications of Digital Signal Processing to Audio and Acoustics
(Kluwer Academic, Boston, 1998)
130. C.-W. Tang, H.-M. Hang, A feature-based robust digital image watermarking scheme. IEEE
Trans. Signal Process. 51(4), 950–959 (2003)
131. Y.Q. Lin, W.H. Abdulla, Multiple scrambling and adaptive synchronization for audio
watermarking, in Digital Watermarking, ed. by Y.Q. Shi, H.-J. Kim, S. Katzenbeisser. Lecture
Notes in Computer Science, vol. 5041 (Springer, Berlin/Heidelberg, 2007), pp. 440–453
132. T. Acharya, A.K. Ray, Image Processing: Principles and Applications (Wiley, Hoboken,
2005)
133. N. Collins, Introduction to Computer Music (Wiley, New York, 2009)
134. A. Lang, Documentation for Stirmark for Audio (2002) [Online], http://amsl-smb.cs.uni-
magdeburg.de/stirmark/doc/index.html
135. H. Zhao, M. Wu, Z.J. Wang, K.J.R. Liu, Nonlinear collusion attacks on independent
fingerprints for multimedia. Proc. ICASSP 5, 664–667 (2003)
136. J.H.L. Hansen, B.L. Pellom, An effective quality evaluation protocol for speech enhancement
algorithms, in Proceedings of International Conference on Spoken Language Processing
(INTERSPEECH), vol. 7, 1998, pp. 2819–2822
137. F. Mustiere, M. Bouchard, M. Bolic, Quality assessment of speech enhanced using particle
filters. Proc. ICASSP 3, 1197–1200 (2007)
138. Y. Hu, P.C. Loizou, Evaluation of objective quality measures for speech enhancement. IEEE
Trans. Audio Speech Lang. Process. 16(1), 229–238 (2008)
139. W.M. Liu, K.A. Jellyman, J.S.D. Mason, N.W.D. Evans, Assessment of objective quality
measures for speech intelligibility estimation. Proc. ICASSP 1, 1225–1228 (2006)
References 199
140. L. Di Persia, M. Yanagida, H.L. Rufiner, D. Milone, Objective quality evaluation in blind
source separation for speech recognition in a real room. Signal Process. 87(8), 1951–1965
(2007)
141. L. Di Persia, D. Milone, H.L. Rufiner, M. Yanagida, Perceptual evaluation of blind source
separation for robust speech recognition. Signal Process. 88(10), 2578–2583 (2008)
142. Y.Q. Lin, W.H. Abdulla, Perceptual evaluation of audio watermarking using objective quality
measures, in Proceedings of ICASSP, 2008, pp. 1745–1748
143. Y. Lin, W. Abdulla, Objective quality measures for perceptual evaluation in digital audio
watermarking. IET - Signal Process. 5(7), 623–631 (2011)
144. T. Rohdenburg, V. Hohmann, B. Kollmeier, Objective perceptual quality measures for the
evaluation of noise reduction schemes, in Proceedings of the 9th International Workshop on
Acoustic Echo and Noise Control (IWAENC), 2005, pp. 169–172
145. SDMI Portable Device Specification, Part 1 (Version 1.0). SDMI (1999) [Online]. http://
ntrg.cs.tcd.ie/undergrad/4ba2.01/group10/technology.html
146. Call for Proposals for Phase II Screening Technology (Version 1.0). SDMI (2000) [Online].
http://ntrg.cs.tcd.ie/undergrad/4ba2.01/group10/technology.html

Audio Watermark Using Matlab

Uploaded by

Copyright:

Available Formats

Audio Watermark Using Matlab

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Audio Watermark Using Matlab

Uploaded by

Copyright:

Available Formats

Yiqing Lin · Waleed H.

Additional material to this book can be downloaded from http://extras.springer.com

ISBN 978-3-319-07973-8 ISBN 978-3-319-07974-5 (eBook)

© Springer International Publishing Switzerland 2015

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Audio watermarking is a technique providing a promising solution to copyrights

Book Motivations and Objectives

Motivated by the booming of the digital media applications, plenty of research

efficient method of evaluating the watermarked audio signals relative to host

Organization of the Book

The chapters in this book are organized as follows.

Auckland, New Zealand Yiqing Lin

2.3 Auditory Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.2.6.2 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6 Perceptual Evaluation Using Objective Quality Measures . . . . . . . . . . . . . 159

A SDMI Standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

B STEP 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

C StirMark for Audio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

D Critical Bandwidth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

E List of Audio Test Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

F Basic Robustness Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

G Nonuniform Subbands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

Fig. 1.1 A generic digital watermarking system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Fig. 2.9 Determination of the critical bandwidth (a) The

Fig. 3.3 Host signal and a watermarked signal by the modified

Fig. 3.14 Host signal and a watermarked signal by

Fig. 5.2 Screenshot of the MATLAB GUI for the MUSHRA

Table 1.1 Classifications of digital watermarking. . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Table 5.13 Results of inter-watermarking Experiment II on

1.1 Information Hiding: Steganography and Watermarking

1.2 Overview of Digital Watermarking

Watermark embedding Watermark detection

Fig. 1.1 A generic digital watermarking system

1.2.1 Framework of the Digital Watermarking System

In general, a digital watermarking system consists of three fundamental parts,

Table 1.1 Classifications of digital watermarking

sw D Embedding .so ; wo ; kw ; ks / (1.1)

where sw should be perceptually similar to so .

we D Detection .sa ; so ; kw ; ks / (1.2)

where we is the extracted watermark. By comparing we with wo , it is verified

1.2.2 Classifications of Digital Watermarking

In terms of different characteristics, digital watermarking can be classified into

1.2.3 Applications of Digital Watermarking

Digital watermarking can be used in a wide range of applications. There is no

1.2.3.1 Copyrights Protection

1.2.3.2 Content Authentication

1.2.3.3 Broadcast Monitoring

1.2.3.4 Copy Control

Most applications of digital watermarking, as discussed so far, have an effect only

1.3 Audio Watermarking for Copyrights Protection

Compared to image and video watermarking, inserting watermark(s) into digital

1.3.1 Requirements for the Audio Watermarking System

Imperceptibility is a prerequisite to practicality. The process of audio watermarking

Robustness is a measure of reliability and refers to the capability of resisting a

Security is a prerequisite to existence. Since the watermarking algorithms are

1.3.1.4 Data Payload

the embedding algorithm. Copyrights protection applications do not require a high

1.3.1.5 Computational Complexity

From a technological point of view, the computational complexity of a watermark-

P .k/ P .k C 1/ and P .k/ P .k 1/ (2.12)