An Acoustic Analysis of Gender Representation in Popular Music

Introduction

This project explores whether large-scale musical data reveal systematic acoustic differences related to artists’ declared gender. While gender inequalities in the music industry are well documented in terms of representation, professional roles, and career trajectories, it remains unclear whether these structural disparities leave detectable traces in the music itself. By combining audio feature analysis, enriched metadata, and statistical modeling, this study provides a descriptive look at possible acoustic patterns across different artist characteristics.

Motivation

Gender inequalities in the music industry

Gender disparities in the music industry are widely documented. According to the USC Annenberg Inclusion Initiative, women accounted for only 37.7% of artists appearing in the 2024 Billboard Hot 100 Year-End Chart. Imbalance is even more pronounced in technical and creative roles: between 2012 and 2017, women represented only 12.3% of songwriters and 2.1% of producers credited on the 600 most popular Billboard songs.

Large-scale metadata studies based on repositories (such as MusicBrainz) confirm that women constitute a minority of identified artists and that gendered patterns are observable across collaboration networks, career trajectories, and musical outputs. These patterns have been linked to structural factors including unequal access to production technologies, professional network segregation, and biased evaluation mechanisms.

However, these inequalities have primarily been studied through sociological and metadata-based approaches focusing on representation and visibility. Whether and how such structural disparities manifest in the acoustic properties of music, remains relatively unexplored. This is an important gap: if differences in access to resources, production environments, and industry networks affect how music is made, we might expect to see systematic differences in the music itself.

Acoustic features reflect the material outcome of musical production. Measures like timbre, pitch distribution, tempo, and loudness, capture aspects of musical expression shaped by compositional choices, production practices, available technology, and stylistic conventions. All of these factors can be influenced by the social and professional conditions under which artists work. Importantly, any acoustic differences found in this study are understood as the result of social, stylistic, technical, and professional processes, not as biological or essential distinctions. In this sense, acoustic analysis is a way of studying cultural artifacts shaped by gendered structures in the music industry.

Research questions

This study has three main goals. First, it establishes empirical baselines for a dimension of musical diversity that has received little attention in computational music analysis. Second, by documenting acoustic variation across artist demographics, it contributes to broader discussions of representation in cultural production. Third, understanding these patterns lays the groundwork for considering how music information systems might interact with or amplify existing disparities.

Modern music recommendation systems rely on collaborative filtering and content-based models that measure similarity using acoustic features. .Research has shown that these systems often reproduce existing inequalities, mainly through popularity bias and feedback loops. If systematic acoustic differences exist across artist characteristics, this raises questions about how content-based representations might interact with recommendation biases.

It is important to note that this project does not evaluate recommendation algorithms or make causal claims. Instead, it offers a descriptive look at acoustic features in a subset of the Million Song Dataset, examining whether meaningful patterns appear when artists are grouped by declared gender. The analysis is exploratory, aiming to document the presence, direction, and magnitude of potential associations.

This analysis is guided by three research questions:

RQ1: Do systematic acoustic differences exist between works associated with male and female artists?

RQ2: Which acoustic dimensions contribute most strongly to any observed differences?

RQ3: Do acoustic patterns vary across musical genres, suggesting genre-specific dynamics?

By focusing on measurable patterns, this study provides a data-driven perspective on gender representation in music while acknowledging the limitations and complexities of both the data and the analytical approach.

Project timeline

Workflow

Week	Task
01.10 - 02.10	Project presentation and topic selection
08.10 - 09.10	Selection of Million Song Dataset subset
15.10 - 16.10	HDF5 file extraction, MusicBrainz API integration
22.10 - 23.11	Week Off
29.10 - 30.10	Metadata enrichment and genre classification
05.11 - 06.11	MVP definition: t-SNE visualization and statistical pipeline
12.11 - 13.11	Midterm presentations
19.11 - 20.11	Development of Dash interactive interface
26.11 - 27.11	RQ1 and RQ2 statistical analyses
03.12 - 04.12	RQ3 ANOVA analyses, sensitivity analyses, representativity assessment
10.12 - 11.12	Results interpretation, Wiki and GitHub documentation
17.12 - 18.12	Final deliverables and presentation

Deliverables and methods

Deliverable

Datasets

Overview of variables in the dataset

The analyses presented in this project rely on a curated subset extracted from the Million Song Dataset (MSD). Although the MSD was originally released as a million song collection, the full dataset is no longer easily accessible in its original form due to its size and age. This study therefore uses a publicly available 10 000 songs subset derived from the original release.

All files in the MSD subset were processed through recursive directory traversal. Each HDF5 file was opened using the h5py library, and metadata from three internal groups (/metadata/songs, /analysis/songs, and /musicbrainz/songs) were extracted using a routine designed to handle missing or heterogeneous fields. In total, 54 attributes were collected per track, including artist identifiers, song-level metadata, and acoustic descriptors such as tempo and loudness. All extracted metadata were compiled into a csv file.

Additional information, including declared artist gender and genre labels, was obtained from the MusicBrainz API. Artist identifiers (artist_mbid) from the Million Song Dataset were used to query the API, following usage rules such as rate limits and user-agent requirements. The field genre_principal was derived from artist tags retrieved via the API. Tags were matched against a predefined mapping that associates common tag keywords with a controlled set of main genres. For each artist, the principal genre was defined as the genre with the highest number of matching tags. Artists for whom no tags matched the mapping were assigned no principal genre.

Segment-level acoustic features were extracted for each track to create compact representations of the music. For every song, the twelve-dimensional arrays segments_pitches and segments_timbre were loaded. To ensure statistical independence, all analyses were done at the artist level. For each artist, segment-level features were averaged across all their tracks for the twelve timbre coefficients and twelve pitch-class profile coefficients. Two global descriptors, loudness and tempo, were also included.

Two finals datasets were created for the analysis. final_song_level.csv contains song-level metadata and acoustic features for interactive visualizations, while final_artist_level.csv aggregates these features at the artist level for statistical analyses. The preprocessing pipeline filters tracks by year, cleans gender and country labels, and computes mean pitch and timbre values at the song and artist levels. The resulting datasets are saved as CSV files for downstream analysis.

Statistical results

The core output of this project is a comprehensive set of statistical analyses, including:

- Univariate tests of gender differences (RQ1), with effect sizes and FDR-adjusted p-values

- Feature importance analysis from Random Forest classification (RQ2)

- Two-way ANOVAs examining genre moderation of gender effects (RQ3)

All analyses were implemented in a fully documented and reproducible codebase: https://github.com/camilledupr/Gender_representation_in_music

Methods

The analytical pipeline followed a three-stage approach: exploratory visual analysis, formal statistical modeling to address the research questions, and global overview through PCA.

Acoustic variables

Loudness (float): Loudness represents the overall perceived intensity of a musical track, expressed in decibels (dB). It provides a global measure of how energetically a piece of music is produced and perceived.

Pitch (segments pitches, 2D array of floats): Pitch is represented through a chroma-based feature (audio representation that describes which musical notes are present in a sound) that encodes the distribution of energy across the twelve pitch classes (notes) of the Western musical scale. The feature is stored as a two-dimensional array, with one value per pitch class for each temporal segment of the track, capturing harmonic and melodic content independently of octave.

Timbre (segments timbre, 2D array of floats): Timbre describes the spectral texture and tone color of the audio signal. It is modeled using a set of coefficients derived from Mel-Frequency Cepstral Coefficients (set of features that describe the short-term spectral shape of an audio signal in a way that approximates human auditory perception), followed by a dimensionality reduction similar to Principal Component Analysis (PCA). These features capture characteristics such as brightness, roughness, and spectral shape, which help distinguish different instruments and vocal qualities.

Tempo (float): Tempo corresponds to the estimated speed of the musical track, expressed in beats per minute (BPM). It reflects the rhythmic pacing of the music and is commonly associated with perceived energy and movement.

Part 1: Exploratory Visual Analysis

Exploratory analyses were conducted to examine the structure of the acoustic feature space, identify potential outliers, and inspect the distribution of metadata variables such as gender and genre. Acoustic features were standardized to zero mean and unit variance using z-score normalization:

Dimensionality reduction was performed using t-distributed Stochastic Neighbor Embedding (t-SNE) with three components, perplexity 30, and PCA initialization to preserve local acoustic similarity patterns. t-SNE minimizes the Kullback-Leibler divergence between probability distributions in high-dimensional (29 dimensions: 12 timbre, 12 pitch, 5 global descriptors) and low-dimensional spaces:

where p_ij represents the similarity between points i and j in the original space, and q_ij represents their similarity in the embedded space. A web-based interface built with the Dash framework enabled interactive exploration of the acoustic space. Users can filter tracks by artist gender, genre, geography, and release year, view metadata for each track, and identify its five nearest neighbors in the embedded space using Euclidean distance. This exploratory component served primarily as a diagnostic and interpretative tool rather than a basis for formal inference. Observations confirmed that genre showed partial clustering, while gender did not form visible clusters, highlighting the need for formal statistical testing.

Part 2: Statistical Modeling

Formal analyses were conducted at the artist level to address the three research questions. Artist-level aggregation ensures statistical independence, as multiple songs by the same artist are not independent observations.

RQ1: Do systematic acoustic differences exist between male and female artists?

To answer this question, we compared the acoustic profiles of male and female artists using Welch's independent-samples t-tests for each of the 26 acoustic variables. Welch's t-test was chosen because it does not assume equal variances between groups, making it more robust when comparing groups of different sizes. The test statistic is:

where X̄_M and X̄_F are the means for male and female artists, s²_M and s²_F are the variances, and n_M and n_F are the sample sizes. The t-statistic measures how many standard errors the two means are apart. A larger absolute value indicates a greater difference between groups. Testing 26 variables simultaneously increases the risk of false positives (Type I error). To control this, we applied the Benjamini–Hochberg False Discovery Rate (FDR) procedure, which adjusts p-values to maintain an expected proportion of false discoveries at 5%. This means that among variables deemed significant, we expect at most 5% to be false positives. Statistical significance (p-value) tells us whether a difference is likely real, but not whether it is meaningful. Cohen's d quantifies the magnitude of the difference:

Interpretation guidelines indicate that | d | ≈ 0.2 represents a small effect (groups overlap substantially), 0.5 represents a medium effect (noticeable difference), and 0.8 represents a large effect (groups are clearly separated).

RQ2: Which acoustic dimensions contribute most to gender differences?

While RQ1 identifies which variables differ significantly, it does not tell us which variables are most important for distinguishing male from female artists when considered simultaneously. To address this, we trained a Random Forest classifier to predict artist gender from all 26 acoustic features. Random Forest is an ensemble method that builds multiple decision trees and combines their predictions through majority voting, reducing overfitting and handling complex interactions between features. The model was configured with 200 trees and maximum depth of 10 to prevent memorization of training data. Balanced class weights compensate for the imbalanced dataset (3.5:1 male-to-female ratio) by giving higher importance to the minority class during training. Z-score standardization ensures all features are on comparable scales. Performance was evaluated using 5-fold stratified cross-validation, which splits the data into 5 parts, training on 4 and testing on 1, while maintaining the original class proportions in each fold. This provides an honest estimate of how well the model generalizes to new data. The primary evaluation metric was balanced accuracy:

Balanced accuracy is the average of recall for each class and is preferred over simple accuracy when classes are imbalanced, as it ensures both classes contribute equally to the evaluation. A balanced accuracy of 0.5 indicates chance-level performance (random guessing), while 1.0 indicates perfect classification. Feature importance was measured using permutation importance, which assesses how much model performance drops when a feature's values are randomly shuffled. If shuffling a feature causes a large drop in accuracy, that feature is important for prediction. By repeating this process 50 times per feature, we obtain stable importance estimates with standard errors. If balanced accuracy is close to 0.5, gender is difficult to predict from acoustic features alone, indicating that gender explains little variance in the acoustic space. Features with high permutation importance are those that, despite overall weak predictive power, contribute most to whatever gender-related signal exists. This helps identify which acoustic dimensions (timbre, pitch, tempo) are most associated with gender differences, even if those differences are small.

RQ3: Do acoustic patterns vary across musical genres?

RQ1 and RQ2 examine average differences across all genres combined, but gender-related patterns might vary by genre due to genre-specific production practices, instrumentation, or vocal styles. To test this hypothesis, we used two-way ANOVA with an interaction term to model the acoustic variable as a function of gender, genre, and their interaction:

where μ is the grand mean, α_i represents the main effect of gender (average difference between male and female artists across all genres), β_j represents the main effect of genre (average differences between genres across both genders), and (αβ)_ij represents the interaction effect (whether gender differences change depending on genre). The ANOVA decomposes total variance into variance explained by gender, genre, and their interaction. The F-statistic tests whether each effect explains significantly more variance than expected by chance. We tested three null hypotheses: that gender has no effect, that genre has no effect, and that gender effects do not vary by genre. Analyses were limited to genres with ≥15 male and ≥5 female artists to ensure sufficient statistical power, resulting in 6 genres (Blues, Country, Jazz, Pop, R&B, Rock) with 477 artists total. If the interaction is significant (p < 0.05), gender differences vary across genres. For example, the male-female gap might be large in Jazz but small in Pop, suggesting that genre-specific factors such as production conventions, instrumentation, and audience expectations modulate how gender manifests acoustically. This would indicate that gender effects are context-dependent rather than universal. If the interaction is not significant (p ≥ 0.05), gender differences are consistent across genres, suggesting that gender-related acoustic patterns reflect broader structural factors such as access to resources and professional networks that operate similarly across musical contexts. If the main effect of genre is much stronger than the gender effect (larger F-statistic), this confirms that genre is a far more important source of acoustic variation than gender, contextualizing RQ1 findings by showing that even when gender differences are statistically significant, they are small relative to genre differences.

Part 3: Global Structure Analysis

Principal Component Analysis (PCA) was applied to the standardized acoustic feature set using a balanced artist sample to assess the overall structure of the acoustic space and the relative importance of gender as a source of variation. PCA identifies orthogonal directions of maximum variance, with the proportion explained by component i computed as:

where λ_i are the eigenvalues of the covariance matrix. Gender differences on PC1 and PC2 were tested using Welch's t-tests. The representativity of the 10,000-track subset was evaluated by comparing summary statistics for key acoustic variables and metadata distributions with those reported in the original Million Song Dataset documentation.

Results

RQ1: Acoustic differences and artist gender

The artist-level univariate analysis, based on Welch's t-tests with False Discovery Rate (FDR) correction, shows clear acoustic differences between male and female artists. Among the 26 acoustic variables kept after preprocessing, twelve show statistically significant differences (P_FDR < 0.05).

Table 1: Acoustic Variables with Significant Gender Differences (Welch's t-test with FDR Correction)

Variable	PFDR	Cohen's d	Direction (Male − Female)
timbre_4_mean	0.000234	−0.330	Female > Male
timbre_3_mean	0.001389	+0.285	Male > Female
timbre_5_mean	0.003506	+0.258	Male > Female
timbre_9_mean	0.003506	+0.251	Male > Female
pitch_9_mean	0.008023	+0.240	Male > Female
pitch_7_mean	0.010586	+0.234	Male > Female
pitch_0_mean	0.023232	+0.211	Male > Female
timbre_0_mean	0.033686	−0.197	Female > Male
timbre_7_mean	0.043102	+0.174	Male > Female
pitch_4_mean	0.045683	+0.178	Male > Female
pitch_11_mean	0.047283	+0.175	Male > Female
pitch_3_mean	0.049764	+0.165	Male > Female

The largest differences are found in timbre-related features, especially timbre_4_mean, timbre_3_mean, and timbre_5_mean, which show small to moderate effects (Cohen's d ≈ 0.26–0.33). Pitch-class profile features also play a role, but with slightly smaller effects. Overall, all significant differences are small to moderate in size (|d| ≈ 0.17–0.33), suggesting considerable overlap between male and female artists despite the statistical significance of the results.

Figure 1. Distribution of timbre_4_mean by artist gender. Box plots show median, interquartile range, and whiskers. Despite statistical significance (P_FDR=0.0002, Cohen's d=−0.33), substantial overlap exists between groups, indicating subtle rather than categorical differences.

RQ2: Acoustic features associated with gender

To identify which acoustic features contribute most to gender differences, a Random Forest classifier was trained at the artist level using all 26 acoustic variables. Model performance was evaluated with 5-fold stratified cross-validation.

The model reached a mean balanced accuracy of 0.543 (± 0.023) and a mean ROC–AUC of 0.686 (± 0.051), indicating performance only slightly better than chance. The low classification performance reflects both the strong class imbalance and the limited separation between male and female artists in the acoustic feature space.

Figure 2. Permutation feature importance with standard errors (50 repetitions). Timbre coefficients dominate discriminative capacity, though overall model performance remains modest (balanced accuracy = 0.543).

Table 2: Acoustic Features with the Highest Importance Scores (Random Forest Classification)

Variable	Importance	Std
timbre_5_mean	0.096	0.012
timbre_4_mean	0.087	0.011
loudness	0.023	0.006
pitch_7_mean	0.023	0.006
timbre_7_mean	0.021	0.006
timbre_3_mean	0.021	0.006
pitch_9_mean	0.016	0.005
timbre_0_mean	0.015	0.005
timbre_1_mean	0.012	0.005
timbre_9_mean	0.012	0.005

Timbre-related features make up eight of the ten most important variables, suggesting that spectral characteristics are the main acoustic dimensions associated with gender-related patterns. Pitch-class features and the global descriptor loudness play a more limited role.

The model's low classification performance, despite clear differences at the group level, points to an important distinction: statistically detectable differences do not imply reliable prediction at the individual level. Gender-related acoustic variation is present, but it is small compared to other sources of variation such as genre, artist identity, and production style.

RQ3: Genre-dependent gender patterns

To examine whether gender-related acoustic differences vary across musical genres, a two-way ANOVA was performed at the artist level on the five most significant variables identified in RQ1. The analysis was limited to genres with sufficient representation (at least 15 male and 5 female artists), resulting in a sample of 530 artists across six genres: Blues, Country, Jazz, Pop, R&B, and Rock.

Table 3: Two-Way ANOVA Results for Genre × Gender Interaction

Variable	Gender F	Gender P	Genre F	Genre P	Interaction F	Interaction P	Significant?
timbre_4_mean	8.22	0.004	31.28	<0.001	0.51	0.768	No
timbre_3_mean	3.33	0.068	2.83	0.016	3.96	0.001	Yes
timbre_5_mean	5.16	0.023	20.15	<0.001	0.88	0.496	No
timbre_9_mean	0.72	0.396	9.24	<0.001	0.98	0.428	No
pitch_9_mean	2.34	0.127	3.98	0.002	1.87	0.099	No

The results show that only one variable (timbre_3_mean) displays a statistically significant Gender × Genre interaction (p = 0.001). For the other variables, gender differences remain largely consistent across genres. In the case of timbre_3_mean, the interaction reflects clear variation in the size of the gender difference between genres, with the largest gaps observed in Jazz, Rock, and Blues. In the case of Pop, the median for female artists is slightly higher than for male artists (visible in box plot below), but the mean difference shows a positive value due to several high-value outliers in the male distribution.

Figure 3. Genre-specific gender patterns for timbre_3_mean. (A) Box plots comparing male and female artists within each genre. (B) Male–female differences per genre. Jazz shows the largest difference but smallest female sample, requiring cautious interpretation.

Overall, the analysis shows limited evidence for genre-specific gender effects. Only one of the five examined acoustic variables displays clear variation in gender differences across genres, indicating that genre moderates gender-related patterns in some cases but not systematically. More broadly, genre emerges as a much stronger source of acoustic variation than gender, shaping musical characteristics to a far greater extent.

Global structure

A Principal Component Analysis (PCA) on a balanced artist sample (n = 223 artists per gender) indicates that the first two components account for 22.2% and 9.6% of the total variance, respectively. Both PC1 and PC2 show statistically significant gender differences (Welch's t-tests), but the scatter plot reveals considerable overlap between male and female artists. This reinforces the conclusion that gender-related acoustic variation is a relatively minor dimension within the overall acoustic space.

Figure 4. Principal Component Analysis projection colored by artist gender (balanced sample: n=223/gender). PC1 (22.2% variance) and PC2 (9.6% variance) show significant gender separation but substantial overlap confirms gender is a minor axis of acoustic differentiation.

Discussion and Limitations

Interpretation of findings

Across all analyses, the results indicate that gender-related acoustic differences do exist, but they are small in magnitude and strongly shaped by musical context.

At the artist level, several acoustic features differ systematically between works associated with male and female artists. About half of the tested variables show statistically significant differences, particularly timbre-related coefficients, with weaker effects for pitch-based features. However, the effect sizes are small to moderate, meaning that the distributions of male- and female-associated music largely overlap. These differences reflect subtle shifts in average tendencies rather than clear-cut acoustic categories.

This subtlety is confirmed by the classification results. Although the Random Forest model performs slightly better than chance, its predictive power remains low. In practice, this means that gender cannot be reliably inferred from acoustic features alone. Still, the model consistently relies most on timbre features, which aligns with the univariate findings. The contrast between statistical significance and weak prediction highlights an important point: group-level differences do not necessarily translate into accurate individual-level classification. In musical data, gender-related signals are present but are easily outweighed by other factors, such as genre, artist identity, and production choices.

Finally, genre plays a moderating but secondary role. Interaction effects between gender and genre are rare, indicating that most gender-related acoustic differences are relatively stable across styles. When moderation does occur, its strength varies by genre, with somewhat larger differences observed in Jazz, Rock, and Country. Nevertheless, genre main effects are consistently much stronger than gender effects, confirming that stylistic context is the primary driver of acoustic variation.

Overall, these findings suggest that gender-related acoustic patterns in music are real but weak. They are best understood as one minor dimension within a much richer acoustic and stylistic landscape, rather than as a defining or predictive characteristic.

Quality assessment

Genre distribution comparison

Several analyses support the overall quality and internal consistency of the dataset and results. The Genre Distribution Comparison plot shows that the 10,000 song subset closely mirrors the full dataset in terms of genre representation, with minor deviations in specific genres like Rock and Pop_Rock. This alignment suggests that the subset is representative of the broader Million Song Dataset (MSD) in terms of genre diversity. The subset data are shown in pink, while the full dataset is shown in blue. When purple appears in certain plots, it indicates that the data point is shared by both datasets.

Timber mean comparison

Pitch mean comparison

The Timbre Mean Comparison and Pitch Mean Comparison plots further validate the subset’s representativeness. While the subset of 10,000 songs is compared to 94,482 songs , the trends in timbre and pitch means are largely consistent between the two. The timbre and pitch features in the subset closely follow the patterns observed in the other dataset, with only minor deviations in specific dimensions ( timbre_1 and pitch_3). This consistency reinforces the reliability of the subset for exploratory acoustic analysis.

Gender distribution comparison

Principal Component Analysis (PCA) results indicate that the first two components explain 20.5% and 9.2% of the total variance, respectively, and both exhibit statistically significant gender differences. However, the strong overlap between gender groups in the PCA space suggests that gender explains only a small fraction of the global acoustic variance. This aligns with the Gender Distribution Comparison plot, which shows a similar distribution of male and female artists in both the subset and the full dataset , further confirming that gender-related acoustic differences are subtle and not definitive.

Top ten country distribution comparison

The Top 10 Countries Distribution Comparison plot reveals that the subset maintains a similar geographic distribution to the full dataset, with the US being the most dominant country in both. This geographic consistency supports the subset’s representativeness and reduces concerns about sampling bias related to regional musical styles.

Limitations

Despite these strengths, several limitations constrain the interpretation and generalizability of the findings. The dataset’s temporal coverage, spanning mostly from the 1970s to the early 2010s, limits its applicability to contemporary music production. Since the early 2010s, changes in production technologies, distribution models, and industry structures (such as the rise of streaming platforms and home recording) may have altered how gender-related patterns manifest in acoustic features. The findings should therefore be understood as characteristic of the historical period represented in the MSD rather than reflective of current musical practices.

Geographic and cultural biases further restrict generalizability. The Top 10 Countries Distribution Comparison plot highlights the dominance of artists from North America (particularly the US) and Western Europe in both the subset and the full dataset. This geographic concentration means the results are most applicable to Western popular music traditions. Musical practices in other cultural contexts may follow different aesthetic norms, production conventions, and gender dynamics, which are not captured in this dataset.

Genre imbalance represents another constraint. The Genre Distribution Comparison plot shows that Rock and Pop are heavily overrepresented, while genres like Hip-Hop, Electronic, and non-Western traditions are underrepresented. Although stratification and weighting strategies were applied, this imbalance limits the ability to draw robust conclusions about underrepresented genres. Genre concentration may also interact with gender distributions, as some genres have historically exhibited pronounced gender imbalances.

Limitations in gender metadata are particularly significant. The Gender Distribution Comparison plot reveals that approximately 54% of tracks in the subset lack gender information, reducing the effective sample size for gender analysis. The binary encoding of gender (Male/Female) does not capture the diversity of gender identities, and the pronounced class imbalance between male and female artists further constrains statistical power and model robustness.

From an inferential standpoint, the study’s descriptive and exploratory design precludes causal claims. While genre was controlled for and analyses were conducted at the artist level, the results document associations rather than causal mechanisms. Multiple explanations may underlie the observed patterns, including unequal access to production resources, gendered differences in musical training, genre segregation driven by industry gatekeeping, disparities in professional networks, and stylistic choices shaped by market positioning. The absence of contextual data on production environments, training backgrounds, and career trajectories limits the ability to evaluate these mechanisms directly.

Finally, the findings have implications for music information systems that must be interpreted cautiously. The weak predictive accuracy and substantial overlap in acoustic space, as seen in the Timbre Mean Comparison and Pitch Mean Comparison plots, suggest that gender-related acoustic variation is dominated by genre, artist style, and production context. Algorithmic biases in recommendation systems are therefore more likely driven by collaborative filtering, historical listening patterns, and feedback loops than by acoustic similarity alone. Nonetheless, even weak acoustic signals may interact with other bias sources or be amplified through iterative recommendation processes, particularly given the central role of timbre features in content-based models.

All interpretations must be framed ethically. The observed acoustic differences are not understood as intrinsic or biologically determined but as outcomes of social, stylistic, technical, and professional processes shaped by structural inequalities in the music industry. The analysis does not support essentialist interpretations of gender and should not be used to justify gender-based categorization or inference. Instead, it highlights the importance of addressing inequality at its structural roots rather than through acoustic proxies.

Credits

Course: Foundation of Digital Humanities (DH-405), EPFL

Professor: Frédéric Kaplan

Supervisor: Alexander Rusnak

Authors: Camille Dupré Tabti, Olivia Robles

Date: December 17, 2025

GitHub Repository: https://github.com/camilledupr/Gender_representation_in_music

Acknowledgments: This project uses a subset of the Million Song Dataset created by Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian Whitman, and Paul Lamere. Gender metadata was enriched through the MusicBrainz API. We thank the computational music analysis community for methodological foundations.