PaperlayerView exact revision on arXiv

Source-linked AI summary

The Evolution of Popular Music: USA 1960-2010

Matthias Mauch, Robert M. MacCallum, Mark Levy, Armand M. Leroi

arXiv:1502.05417v1physics.soc-phcs.SD

TL;DR

Claims about pop history have often been anecdotal, and earlier quantitative studies have been limited in musical focus or sample size. The authors analyze Billboard Hot 100 recordings with audio features and text-mining methods, finding changing musical diversity and evolution marked by periods of stasis and rapid change.

Abstract

from arXiv · show

In modern societies, cultural change seems ceaseless. The flux of fashion is especially obvious for popular music. While much has been written about the origin and evolution of pop, most claims about its history are anecdotal rather than scientific in nature. To rectify this we investigate the US Billboard Hot 100 between 1960 and 2010. Using Music Information Retrieval (MIR) and text-mining tools we analyse the musical properties of ~17,000 recordings that appeared in the charts and demonstrate quantitative trends in their harmonic and timbral properties. We then use these properties to produce an audio-based classification of musical styles and study the evolution of musical diversity and disparity, testing, and rejecting, several classical theories of cultural change. Finally, we investigate whether pop musical evolution has been gradual or punctuated. We show that, although pop music has evolved continuously, it did so with particular rapidity during three stylistic "revolutions" around 1964, 1983 and 1991. We conclude by discussing how our study points the way to a quantitative science of cultural change.

Introduction

The study addresses the lack of rigorous, quantitative tests of claims about popular music’s history by asking how its variety changed and whether its evolution was continuous. It combines a clearly defined chart-based sample with musically meaningful audio features to examine those questions.

Introduction

  • The study asks whether popular music’s variety increased or decreased and whether its evolution was continuous or marked by discontinuities.The authors frame these as questions that can be tested using quantitative data and statistics rather than anecdote alone.Source: Introduction · p. 1
  • The authors combine a large, clearly defined diachronic dataset with detailed analysis of musically meaningful audio features.They contrast this approach with earlier studies that lacked scientific rigor, focused on technical audio properties, or were hampered by sample size.Source: Introduction · p. 1
  • The sample contains 17,094 30-second excerpts from songs on the US Billboard Hot 100 between 1960 and 2010, covering 86% of the chart.The sample has a small bias toward missing songs in earlier years.Source: Introduction · p. 1
  • The analysis converts harmonic and timbral audio features into 16 LDA Topics, representing each song as a distribution over those Topics.The harmonic Topics capture classes of chord changes, while the timbral Topics capture timbres linked to expert annotations.Source: Introduction · p. 3

Results

The Billboard Hot 100 evolved continuously, but change accelerated around three revolutions and reflected shifts in musical Topics and Styles. Diversity rebounded after an early-1980s low, while the British Invasion amplified—but did not initiate—trends already underway.

The Evolution of Topics

  • H8’s major-chord pattern was the most common harmonic Topic in 43 of 50 years, while H1’s dominant-seventh frequency fell about 75% from 1960 to 2009.H3 more than doubled from 1967 to 1977; H6 rose from 1978 to 1985 before declining, while chord-sparse H5 rose rapidly to a 1993 peak with hip hop and rap.Source: The Evolution of Topics · pp. 3–4
  • Timbral Topics showed recurring fashions: T4 appeared sinusoidal, T5 completed two cycles with peaks in 1966 and 1985, and T1 rose until 1990 before declining.T3 shared H5’s dynamics and association with hip-hop-related genres; T5’s larger peak coincided with H6’s peak.Source: The Evolution of Topics · p. 4

The varieties of music

  • The authors classified songs into 13 Styles by clustering principal components of Topic frequencies; Last.fm tag enrichment linked the clusters to consumer-perceived musical structure.Rap-enriched Style 2 was rare before the mid-1980s, then expanded to become the largest Style for the next thirty years before contracting in the late 2000s.Source: The varieties of music · p. 4

Musical diversity has not declined

  • Topic diversity and disparity fell to a minimum around 1984, then rebounded to early-2000s maxima; their dynamics were not explained by song counts or recent-song overrepresentation.The early-1980s decline was associated with dominant timbral Topics T1 and T5; recovery accompanied their relative decrease and an increase in T3, alongside shifts in Styles.Source: Musical diversity has not declined · p. 6

Musical evolution is punctuated by revolutions

  • 4- to 6-fold faster change from peak to succeeding trough marked revolutions around 1964, 1983, and 1991, amid otherwise ceaseless evolution.The 1991 revolution was the largest of the three.Source: Figure 5 · p. 7, Musical evolution is punctuated by… · p. 8
  • The revolutions coincided with expansions and contractions of distinct Styles: rap-enriched Style 2 expanded in 1991 as rock-enriched Styles 5 and 13 contracted.In 1964, soul- and rock-enriched Styles expanded at the expense of doowop-enriched Styles; in 1983, new wave, disco, and hard rock-enriched Styles expanded as soft rock, country, and soul/R&B-enriched Styles contracted.Source: Musical evolution is punctuated by… · p. 8

The British did not start the American revolution of ’64

  • The four principal components’ downward trajectories began before the British Invasion, so British acts could not have been solely responsible for the 1964 changes.The Beatles and Rolling Stones were ahead of chart trends on PC1 and PC2, but resembled the wider Hot 100 on PC3 and PC4.Source: The British did not start the American… · p. 8

FIGURE 6

The figure tracks four principal components around the 1964 revolution, comparing chart-wide trajectories with the Beatles and Rolling Stones. The bands had lower PC1 and PC2 medians than other artists, but the population trends predated the British Invasion.

FIGURE 6

  • All four population PC medians were already decreasing before the British Invasion began in late 1963, so its acts could not be solely responsible for the changes.Panel A shows quarterly medians in the six years flanking the 1963–Q4 revolution peak.Source: Figure 6 · p. 9
  • The Beatles and Rolling Stones had significantly lower PC1 and PC2 medians than other artists, but not PC3 or PC4.PC1 reflects increased major-chord use and decreased bright speech; PC2 reflects increased guitar-driven aggression and decreased mellow vocals.Source: Figure 6 · p. 9
  • The two bands’ PC1 and PC2 values anticipated the Hot 100’s trajectory, while their PC3 and PC4 values were on-trend.Together, the findings suggest the bands exploited the revolution and may have fanned its flames through imitation by other artists.Source: Figure 6 · p. 9, FIGURE 6 · p. 10

Discussion and Conclusions

The study offers a quantitative account of US popular music while noting limits in its features, chart sample, and causal explanations. Its statistical tools and digital corpora may support evolutionary analysis across modern culture.

Discussion and Conclusions

  • The study provides a quantitative picture of US popular music over fifty years and points toward a scientific study of musical change.Because the approach is audio-based, it can also be applied to music without scores, including pre-Modern cultures.Source: Discussion and Conclusions · p. 10
  • The findings cover only part of song complexity, and other features may yield different results.Listener genre-tags support the classifications, but the measures capture only a fraction of the phenotypic complexity of even a simple song.Source: Discussion and Conclusions · p. 10
  • Because the Hot 100 is a biased subset of US releases, the conclusions cannot be extended to all releases.The study sampled only a small fraction of new singles released from 1960 to 2010.Source: Discussion and Conclusions · p. 10
  • The study identifies musical dynamics but does not address their causes.A causal account would need to explain inheritance, musical novelty, and constraints; inheritance and its constraints are described as obscure.Source: Discussion and Conclusions · p. 10
  • Latent Dirichlet Allocation and Foote Novelty can analyze evolving assemblages and rate discontinuities in distance sequences, respectively.The passage presents these tools as applicable to many kinds of assemblages and phenotypes.Source: Discussion and Conclusions · p. 10
  • Statistical tools and large digital corpora may permit evolutionary analysis of many dimensions of modern culture and help constrain and inspire cultural-evolution theories.The passage compares this anticipated role to that of the fossil record in the study of life.Source: Discussion and Conclusions · p. 11

Authors’ contribution

The passage assigns data provision, analysis, study design, coordination, and manuscript writing among the authors.

Authors’ contribution

  • ML provided data; MM, RMM, and AML analyzed it, while MM and AML conceived, designed, coordinated, and wrote the study.All authors gave final approval for publication.Source: Authors’ contribution · p. 11

Funding

The passage identifies Matthias Mauch's fellowship funding.

Funding

  • Matthias Mauch was funded by a Royal Academy of Engineering Research Fellowship.Source: Funding · p. 11

Supporting Information

The passage lists four authors.

Supporting Information

M.1 The origin of the songs

The methods link chart entries to audio recordings, extract harmonic and timbral features, and convert them into discrete lexica for song-level analysis.

M.1 The origin of the songs

  • 17,094 30-second recordings cover 86% of weekly Hot 100 charts; volunteer checks found 3.8% poor audio and no significant topic-mean shifts after excluding poor tracks.Coverage was complete for 1960–2009; poor-quality recordings were more common in the 1960s, but removing them did not significantly change estimated topic means for any year in that decade.Source: Figure M1 · p. 16, §M.1 · p. 16

M.2 Measuring Harmony.

  • Chroma features are 12-dimensional pitch-class profiles representing activation strength across the 12 chromatic-scale pitch classes; NNLS Chroma is extracted at the timbre-feature frame rate (1024-sample step, 23 ms; 43 per second).Source: §M.2 · p. 16

M.3 Measuring Timbre.

  • Timbre is represented by 12 MFCCs, one delta-MFCC, and one zero-crossing-count feature per frame.The delta-MFCC is the difference between consecutive values of the 0th MFCC component; zero-crossing count captures sign changes between consecutive samples.Source: §M.3 · p. 17

M.4 Making musical lexica

  • PCA and Gaussian-mixture clustering discretize timbre into 35 classes, while harmony uses 192 chord changes plus an ambiguous 193rd NA label; songs become class-count vectors.The harmonic lexicon describes chord changes using relative root positions and chord types, making it key-independent.Source: §M.4 · pp. 17–18

M.5 Semantic lexicon annotation

The supplementary methods annotate sound classes, model song-level topic mixtures, interpret those topics, and use them to identify styles in the chart recordings.

M.5 Semantic lexicon annotation

  • Ten annotators labeled 35 timbre classes from concatenated clips; the top label received mean agreement from 7.5 of 10 annotators.The annotators chose five descriptors per class from a controlled vocabulary of 34 labels.Source: §M.5 · pp. 18–19

M.6 Topic extraction

  • LDA was fitted separately to timbre and harmony counts, estimating each song’s topic mixture and each topic’s lexicon distribution; eight topics were set for each model.The song-topic probabilities can be interpreted as the proportion of frames in a song belonging to each topic.Source: §M.6 · pp. 19–20

M.7 Semantic topic annotations

  • T-topics inherit weighted human sound labels, while H-topics are interpreted from their probable chord changes; Table 1 reports the topic interpretations.For example, H-Topic 4 is characterized by a major chord followed by another major chord seven semitones higher.Source: §M.7 · p. 20, Table 1 · p. 21

M.8 User-generated tags

  • Last.fm tags, usually genre-related, were available for 94% of songs, with a mean of 2.7 tags per song.Artist references were removed and semantically identical tags were combined.Source: §M.8 · p. 21

M.9 Identifying musical Styles clusters: k-means and silhouette scores

  • 13 audio-derived style clusters were selected by mean silhouette width after PCA reduced the 16 topic probabilities to 14 components.The analysis compared cluster counts from 2 to 25, using 50 random initializations for each count.Source: §M.9 · p. 21, Figure M4 · p. 22

M.10 Diversity metrics

The section defines chart diversity using song counts, effective numbers of styles and topics, and disparity. Topic diversity declined significantly in the 1980s mainly because timbral-topic diversity fell, while harmonic-topic diversity showed no sustained decline.

M.10 Diversity metrics

  • Song count is the simplest complexity measure and is used to show whether other diversity measures are intrinsically linked to it.Source: §M.10 · p. 22
  • Style diversity is the effective number of styles, calculated yearly from the song proportions across 13 styles.The measure ranges from 1 to 13.Source: §M.10 · pp. 22–23
  • Yearly topic diversity uses effective topic counts calculated from the proportions of audio frames assigned to each topic across a year's songs.The overall measure averages the harmonic- and timbral-topic diversity measures.Source: §M.10 · p. 23
  • Timbral-topic diversity drove the significant 1980s decline in topic diversity, while harmonic-topic diversity showed no sustained decline.The section defines overall topic diversity as the average of harmonic- and timbral-topic diversity.Source: §M.10 · p. 23
  • Disparity is measured as total standard deviation across the 14-dimensional principal-component representation of the topics.The measure is the square root of total variance.Source: §M.10 · p. 23

M.11 Identifying musical revolutions

Foote novelty detection on quarterly distances identifies three major musical change regions, with significant long-term change but none at scales under two years. The section also describes how styles and associated tags are examined around the revolutions.

M.11 Identifying musical revolutions

  • Foote novelty detection convolves a checkerboard kernel along a pairwise-distance matrix of quarterly mean principal components; transitions between homogeneous eras yield high scores.The quarters are formed by pooling the 14-dimensional principal-component data by songs' first chart-entry quarter.Source: §M.11 · pp. 23–24
  • Three major novelty peaks appear in the early 1960s, early 1980s, and early 1990s across a wide range of kernel sizes.Using a 12-quarter half-width, the three change regions are significant at p < 0.01; no changes are significant below two years, while all quarters show significant change at large scales.Source: §M.11 · p. 24

M.12 Identifying Styles that change around each revolution

  • Style frequencies across the 24 quarters flanking each peak are modeled quadratically to estimate annual change, and tag enrichment identifies associated genre tags.Table S2 covers styles changing around each revolution and their over-represented tags in those quarters.Source: §M.12 · p. 24, Table S2 · p. 27

S Supplementary Text & Tables

  • Supplementary material lists eight harmonic-pattern groups and eight tag triplets, including “drums, aggressive, percussive” and “calm, quiet, mellow.”Source: S Supplementary Text & Tables · p. 25
Loading 1502.05417v1…