Evolution of The Informational Complexity of Contemporary Western Music
Evolution of The Informational Complexity of Contemporary Western Music
Evolution of The Informational Complexity of Contemporary Western Music
175
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
200
2.2 Codewords X
H(Y |X) = p(x)H(Y |X = x) =
To estimate complexity, we defined “codewords” for each xX (1)
song across four dimensions (pitch, loudness, timbre, and −
X
p(x)
X
p(y|x) log2 p(y|x)
rhythm), similar to a previous study [26]. Each codeword xX yY
is based on a segment of the song. Pitch and timbre code-
words are vectors containing the pitches (based on the bi- where X and Y are possible codewords, p(x) is the
nary presence of each of 12 pitches in the chromatic scale) probability of observing codeword x and p(y|x) is the
and timbres (based on analysis of the audio signal, with probability of observing codeword y given the previous
11 components thresholded into three bins) present in the codeword x.
segment. Loudness codewords are equal to the binned
maximum decibel value of the segment. Similarly, rhythm 1 Code is available at https://github.com/tjparmer/music-complexity.
176
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
177
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
Sample
0.09 0.09 0.09 Hot 0.09
100
0.06 0.06 0.06 0.06
0.03 0.03 0.03 0.03
0.00 1 2 3 4 5 0.00 1 2 3 4 5 0.00 0.1 0.4 0.7 1.0 1.3 0.00 1 2 3 4 5
Complexity (bits) Complexity (bits) Complexity (bits) Complexity (bits)
Hot 100 Hot 100 Hot 100 Hot 100
all songs all songs all songs all songs
pop pop pop pop
hip hop hip hop hip hop hip hop
jazz jazz jazz jazz
heavy metal heavy metal heavy metal heavy metal
electronica electronica electronica electronica
0.1 0.2 0.3 0.1 0.2 0.3 0.0 0.1 0.2 0.1 0.2 0.3 0.4
Variance Variance Variance Variance
Figure 2. Feature complexities and variances. Complexity distributions are shown (in bins of 0.1 bits, except for timbre
which is in bins of 0.02 bits). The variance plots include 95% confidence intervals in black (although confidence intervals
are smaller than the symbol and not visible), based on 1000 bootstrap samples of 1000 songs from the respective genre.
on the charts do not sound too similar to their contempo- they prefer pitch and loudness complexity, they may prefer
raries but rather maintain a degree of uniqueness which is hip hop or jazz. If they care about timbre and rhythm over
statistically consistent over time. pitch and loudness, they may prefer heavy metal. There is
a positive correlation between pitch and loudness complex-
ity (Pearson’s r=0.77) across all songs, suggesting that gen-
3.4 Complexity Across Genres
res tend to have high pitch and loudness complexity (e.g.
Let us turn our attention to musical genres and their com- hip hop, jazz) or low pitch and loudness complexity (e.g.
plexity. As some genres may be characterized by complex heavy metal). There is also a negative correlation seen be-
harmonic structures or simple, repeated patterns, we ex- tween timbre and rhythm complexity (Pearson’s r=-0.55),
pect to see differences across different genres in terms of suggesting that rhythmic complexity decreases with higher
complexity. For example, jazz is often considered to have timbre complexity (although this is not true for metal gen-
complex patterns whereas dance music may be assumed to res).
use simpler rhythmic patterns. Our measurement concurs Interestingly, the Hot 100 is similar to the pop genre in
with such speculation, but finds that different subsets of feature means and variances (although statistically differ-
genres may be relatively complex across one or two fea- ent). Both pop music (whose songs are given no genre-
tures but not others. For instance, electronic and dance specific term with higher weight than ‘pop’) and the Hot
styles tend to have high pitch complexity values, whereas 100 (whose songs are primarily classified as genres other
jazz and blues have high loudness complexity values. The than ‘pop’) have near average values of pitch, loudness,
highest timbre complexity values belong to electronic gen- and rhythm complexity, and lower than average values of
res, although metal also scores highly, but electronic gen- timbre complexity, while also having smaller variance than
res have reduced rhythmic complexity which is instead the other selected genres (refer to Figures 2 and 6). This
maximal in jazz, progressive and vocal genres. may suggest that listeners expect the same from listening
We found that a variety of common genres were sig- to the Hot 100 as they do when listening to music labeled
nificantly different from a random sample drawn from the as ‘pop’: mildly surprising songs that do not vary too much
overall distribution in terms of each feature complexity in complexity and which are sonically predictable.
(based on a two-sample Kolmogorov-Smirnoff nonpara- One may also expect similar genres to share similar
metric test as well as 95% confidence intervals of the complexity scores. We used agglomerative clustering on
means of each feature), with the exception that pop was not genres represented with over 5000 songs in the dataset (a
rhythmically distinct. Thus each genre seems to have dis- total of 41 genres), using Euclidean distance between the
tinctive complexity features that describe its songs: jazz is genre mean complexity scores of each feature, and then
relatively complex (except in terms of timbre), hip hop has used the silhouette coefficient [24, 28] to find nine opti-
higher than average pitch and loudness complexity, heavy mal communities. The result matches intuitive expecta-
metal has high rhythm complexity but low pitch and loud- tions, such that rock genres are grouped together as are
ness complexity, and electronica has high timbre complex- electronic genres; interestingly jazz is grouped with hip
ity but low rhythm complexity (Fig. 6). hop and rap, due to these genres having similar complexity
This pattern may be indicative of some trade-offs that scores. These results suggest that a genre may be defined,
listeners make. If they prefer timbre at the expense of to some degree, by its pitch, loudness, timbre, and rhyth-
rhythmic complexity, they may prefer electronic genres. If mic complexity.
178
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
0.3
Timbre 0.2
0.1
3.0
Rhythm 2.8
2.6
1960 1970 1980 1990 2000 2010
1964 1983 1991
Year
All Songs Hot 100 All Songs best fit Hot 100 best fit
Figure 3. Average complexity over time across features. Linear trend lines (obtained using OLS linear regression) are
shown for each epoch, along with 95% confidence intervals of the mean. Light blue lines indicate the musical revolutions
found in [15].
1.1
Loudness 0.8
0.5
7.8
Timbre 7.3
6.8
0.6
Rhythm 0.5
0.4
1960 1970 1980 1990 2000 2010
1964 1983 1991
Year
Figure 4. Yearly KL divergence of Hot 100 songs across features. Complexity variance, as measured by KL divergence
between a song’s codewords and codewords from other songs released in that year, shows changing trends at the time of
certain musical ‘revolutions’ in 1964, 1983, and 1991 (indicated by light blue lines). Linear trend lines (obtained using
OLS linear regression) are shown for each epoch, along with 95% confidence intervals of the mean.
179
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
0.7
0.6
0.5
Distance
0.4
0.3
0.2
0.1
0.0
ssi ho e
rd z
have house
eu l-out
tr e
fut lectce
ure ro
trip jazz
ch hop
nc op
d p
po isco
sm hip h p
ga oth op
st zz
roo azz f rap
ts unk
ga nceh e
alt heavge r ll
ern y m ock
ea tive m tal
rim ccmd
co liste tal
sofy rocg
blu t ro k
es ck
fol -rock
an k
allon
podie rock
we ock
bri rt ro p
po tish p ck
st- op
ern nge
ve k
po rock
vo wavk
ca e
p
ha l jaz
ra a
chk roc
ati roc
w c
gre eep ous
rd us
da egga
o
tr in
a po
bo
ng ja
ne p ro
e an
da op
ep
s
pr
a e
sy e
un n
in tal r
gru
j er
r
il
pro d ch h
b
r
en
o
te
alt
pe
ex
Figure 5. Dendrogram of top genres. Clustering is based on the Euclidean distance between complexity features. Colors
indicate different communities.
Complexity per genre 100 are distributed around the overall feature means with
0.3 jazz electronica
small variance in complexity, the exception being timbre
Deviation from average
hip hop which is lower than average. Indeed, the Hot 100 is simi-
0.2 lar to songs labeled as ‘pop’, in that pop also has average
0.1 pitch, loudness, and rhythm complexity and low variance.
This result provides evidence of a global, inverted U-
0.0 shaped relationship between popularity and complexity,
0.1 where popular songs are, on average, the most pleasant to
the population. Listeners may expect popular songs to be
0.2 pop
mildly complex, but not to deviate far from expected tim-
Hot 100
0.3 heavy metal
bre or complexity norms. Complexity of the Hot 100 has
in fact been consistent over fifty years in pitch and tim-
Pitch Loudness Timbre Rhythm bre, while increasing recently in rhythm and decreasing in
loudness. Similarly, popular songs continue to maintain
a consistent level of differentiation from their contempo-
Figure 6. Complexity of select genres across features. For raries in terms of pitch, timbre, and rhythm.
each feature, the deviation of the genre complexity mean Certain genres do differ significantly across complex-
from the mean of 100 bootstrapped samples of the same ity features, suggesting that they have specific complexity
size as the genre is shown. Note the similarity between the profiles that help define them. We hypothesize that certain
Hot 100 and the pop genre. genres may ‘make up’ for lack of complexity in one area
by increased complexity in another. Perhaps fans of elec-
tronic genres prioritize complexity in timbre but rhythmi-
4. DISCUSSION
cally simple dance beats, or metal fans prioritize rhythmic
Understanding and characterizing the complexity of music complexity and high volume at the expense of loudness
is an important area of study with both cultural and eco- complexity.
nomic significance. Although music may seem compli- More research needs to be done to fully elucidate the
cated, songs quickly become predictable as you take pre- relationship between complexity and musical appreciation.
vious knowledge into account. This suggests that condi- For example, future research can relate musical complex-
tional entropy may be a useful way to characterize musi- ity to the listening habits of people on a large scale to de-
cal complexity, although our approach here assumes that termine a more fine-grained measure of song popularity.
the uniform distribution of codewords is the state of maxi- The consistency of popular songs over time suggests that,
mum uncertainty and expectations are made based on only collectively, people tend towards songs that are a certain
one previous symbol, which cannot distinguish counts of optimal level of complexity rather than being too simple
repeating patterns or phrasing [1]. Our approach is thus in- or complicated. However, it remains an open question to
trinsic to the song itself and ignores any a priori contextual what degree complexity plays a role in people’s cognitive
information. appreciation of music.
Using this measure, we find that pitch complexity has
been generally stable over the period from 1960 to 2010,
5. REFERENCES
while loudness and rhythm complexity have decreased and
timbre complexity has increased. Complexity norms seem [1] Samer Abdallah and Mark Plumbley. Information dy-
to constrain the most popular songs, as those on the Hot namics: patterns of expectation and surprise in the per-
180
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
ception of music. Connection Science, 21(2-3):89–117, [14] Solomon Kullback and Richard A Leibler. On informa-
2009. tion and sufficiency. The annals of mathematical statis-
tics, 22(1):79–86, 1951.
[2] Noah Askin and Michael Mauskapf. What makes pop-
ular culture popular? Product features and optimal dif- [15] Matthias Mauch, Robert M MacCallum, Mark Levy,
ferentiation in music. American Sociological Review, and Armand M Leroi. The evolution of popular mu-
82(5):910–944, 2017. sic: USA 1960–2010. Royal Society open science,
2(5):150081, 2015.
[3] Daniel E Berlyne. Novelty, complexity, and hedo-
nic value. Perception & Psychophysics, 8(5):279–286, [16] David Meredith. Music analysis and kolmogorov
1970. complexity. In Proceedings of the 19th Colloquio
d’Informatica Musicale (XIX CIM), 2012.
[4] Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian
Whitman, and Paul Lamere. The million song dataset. [17] Leonard B Meyer. Music, the Arts and Ideas: Patterns
In Proceedings of the 12th International Conference on and Predictions in Twentieth-century Culture. Univer-
Music Information Retrieval (ISMIR 2011), pages 591– sity of Chicago Press, 1967.
596, 2011.
[18] Eugene Narmour. The analysis and cognition of basic
[5] Rudi Cilibrasi, Paul Vitányi, and Ronald de Wolf. Al- melodic structures: The implication-realization model.
gorithmic clustering of music based on string compres- University of Chicago Press, 1990.
sion. Computer Music Journal, 28(4):49–67, 2004.
[19] Adrian C North and David J Hargreaves. Experimental
[6] Sander Dieleman, Philémon Brakel, and Benjamin aesthetics and everyday music listening. In The social
Schrauwen. Audio-based music classification with a psychology of music, pages 84–103. Oxford University
pretrained convolutional network. In Proceedings of Press, Oxford, 1997.
the 12th International Conference on Music Informa- [20] Adrian C North and David J Hargreaves. Liking,
tion Retrieval (ISMIR 2011), pages 669–674, 2011. arousal potential, and the emotions expressed by mu-
[7] Tuomas Eerola and Adrian C North. Expectancy-based sic. Scandinavian journal of psychology, 38(1):45–53,
model of melodic complexity. In Proceedings of the 1997.
Sixth International Conference on Music Perception [21] Mark G Orr and Stellan Ohlsson. Relationship between
and Cognition, 2000. complexity and liking as a function of expertise. Music
[8] Peter Foster, Simon Dixon, and Anssi Klapuri. Iden- Perception: An Interdisciplinary Journal, 22(4):583–
tifying cover songs using information-theoretic mea- 611, 2005.
sures of similarity. IEEE/ACM Transactions on Audio, [22] Robert Mitchell Parry. Musical complexity and top 40
Speech, and Language Processing, 23(6):993–1005, chart performance. Technical report, Georgia Institute
2015. of Technology, 2004.
[9] Peter Foster, Anssi Klapuri, and Simon Dixon. A [23] Jeffrey Pressing. Cognitive complexity and the struc-
method for identifying repetition structure in musical ture of musical patterns. In Proceedings of the 4th Con-
audio based on time series prediction. In 2012 Pro- ference of the Australasian Cognitive Science Society,
ceedings of the 20th European Signal Processing Con- 1999.
ference (EUSIPCO), pages 1299–1303. IEEE, 2012.
[24] Peter J Rousseeuw. Silhouettes: a graphical aid to the
[10] Barbra Gregory. Entropy and complexity in music: interpretation and validation of cluster analysis. Jour-
some examples. PhD thesis, University of North Car- nal of computational and applied mathematics, 20:53–
olina at Chapel Hill, 2005. 65, 1987.
[11] Ronald G Heyduk. Rated preference for musical com- [25] Eric D Scheirer, Richard B Watson, and Barry L Ver-
positions as it relates to complexity and exposure coe. On the perceived complexity of short musical seg-
frequency. Perception & Psychophysics, 17(1):84–90, ments. In Proceedings of the 2000 International Con-
1975. ference on Music Perception and Cognition, 2000.
[12] David Brian Huron. Sweet anticipation: Music and the [26] Joan Serrà, Álvaro Corral, Marián Boguñá, Martín
psychology of expectation. MIT press, 2006. Haro, and Josep Ll Arcos. Measuring the evolution
of contemporary western popular music. Scientific re-
[13] Heather D Jennings, Plamen Ch Ivanov, Allan de M ports, 2:521, 2012.
Martins, PC da Silva, and GM Viswanathan. Variance
fluctuations in nonstationary time series: a comparative [27] Ilya Shmulevich and D-J Povel. Measures of temporal
study of music genres. Physica A: Statistical Mechan- pattern complexity. Journal of New Music Research,
ics and its Applications, 336(3-4):585–594, 2004. 29(1):61–69, 2000.
181
Proceedings of the 20th ISMIR Conference, Delft, Netherlands, November 4-8, 2019
182