GlobalDISCO: Auditing Bias in AI Music

• GlobalDISCO is a globally balanced dataset that audits AI-generated music bias by covering 147 languages, 79 countries, and diverse musical genres.
• It systematically compares 73,792 generated tracks with 92,859 country- and genre-specific reference tracks to reveal regional disparities.
• The framework uses embedding models and metrics like FAD and KAD to evaluate mainstream drift and misalignment in text-to-music systems.

Searching arXiv for the GlobalDISCO paper and closely related context. GlobalDISCO is a large-scale, globally balanced dataset designed specifically to audit and benchmark biases in AI-generated music across languages, countries, regions, and genres. It pairs 73,792 AI-generated tracks with links to 92,859 reference tracks from LAION-DISCO-12M, spans 147 languages, 79 countries, and five continents, and was introduced in “Bias beyond Borders: Global Inequalities in AI-Generated Music” (Solak et al., 2 Oct 2025). Built from nearly 10k artist profiles, with 9,300 retained after filtering, it provides a structured benchmark for measuring how well current state-of-the-art commercial text-to-music systems capture worldwide musical diversity and where they systematically fail.

1. Scope and rationale

GlobalDISCO was motivated by a specific gap in the evaluation of generative music systems: research on biases across countries, languages, cultures, and musical genres had remained underexplored, in part because globally representative benchmarks were absent. The dataset was designed to fill that gap by covering 147 languages and 79 countries, which the paper characterizes as an order of magnitude more diverse than prior synthetic music datasets, while directly comparing generated outputs to country- and genre-specific reference distributions (Solak et al., 2 Oct 2025).

The benchmark is geographically balanced by design. It includes mainstream global genres such as pop, rock, hip hop, and k-pop, but also geographically niche or regional styles such as soukous, ghazal, pub rock, and opera. At least 75 generated tracks are provided per country, with a median of 502 and a maximum of 2,861. This balance is central to the dataset’s intended use: not merely cataloguing prompts and outputs, but enabling region- and genre-conditioned auditing under controlled coverage constraints.

The benchmark targets state-of-the-art commercial text-to-music systems rather than open-weight models alone. Four commercial black-box systems are used: Suno v4, Udio v1.5 Allegro, Mureka v6, and Riffusion (FUZZ 0.8). Hyperparameters beyond model version are not exposed by the platforms, so default generation settings are used. This design choice makes GlobalDISCO a benchmark for deployed generative systems as they are actually encountered by end users, while also constraining the degree of low-level experimental control.

2. Construction pipeline

The construction pipeline begins with artist curation from MusicBrainz. Starting from 148k MusicBrainz artist entries with geographic area and linked biography pages, the pipeline augments missing Wikipedia links via English Wikipedia searches and retains only matches consistent in name and at least two attributes such as area or genre (Solak et al., 2 Oct 2025). Instrumental-only artists are excluded in order to focus on multilingual vocal content, whereas artists with mixed metadata, such as classical plus vocal genres, are retained.

Reference linkage is performed against LAION-DISCO-12M by matching artist and channel names and verifying discography overlap when necessary. For each retained artist, the top 10 most viewed LAION-DISCO tracks are kept as references. These references serve as country- and genre-conditioned target distributions against which generated outputs are compared. The benchmark therefore does not treat evaluation as purely prompt-following assessment; it anchors evaluation to empirical musical distributions associated with artists and regions.

Global balancing is enforced algorithmically. From 34k artists satisfying the vocal criterion, up to $t = 374$ artists per country are selected, with $t$ found by binary search to yield roughly 10k artists overall. After filtering for generation success and reference availability, 9.3k artists remain, with at least 10 artists per country. This retained set then supports prompt creation and generation at benchmark scale.

Style prompts and lyrics are derived from artist profiles assembled from MusicBrainz and Wikipedia. These profiles include name, country, genres, instruments, and biographical snippets, and they are fed to Gemini to produce musical style descriptions and synthetic lyrics. Following prior methodology, up to three real lyric samples are used in few-shot prompting where such samples exist; where they do not, lyrics are generated from the profile. A plausible implication is that prompt construction is deliberately standardized while still preserving country- and artist-specific stylistic context.

3. Coverage, stratification, and metadata schema

GlobalDISCO organizes its audit space through both geographic and genre stratification. Regions follow the UN M49 standard, and analyses group countries into 13 world sub-regions (Solak et al., 2 Oct 2025). The paper operationalizes “high-resource” and “low-resource” at the region level as a proxy for availability and coverage in online music and training data. Northern America is treated as high-resource, while Northern Africa, Sub-Saharan Africa, Southern Asia, and Western Asia are treated as lower-resource in the reported analyses. The paper explicitly notes that this distinction is not defined by a numeric threshold.

Genre classes are divided into mainstream and geographic niche categories. Mainstream genres are the top ten most frequent genres in GlobalDISCO. Geographic niche, or regional, genres are selected by a tf-idf-like country-specific score:

$$\text{score} = (\text{relative frequency of genre within country}) \times (\text{inverse of the number of countries in which the genre appears}).$$

For each country, the top-scoring genre is included if it is associated with at least 10 artists. This procedure yields regional styles including soukous, pub rock, opera, and ghazal.

The dataset metadata reflect both artist provenance and generation context. GlobalDISCO provides generated audio and paired links to LAION-DISCO-12M references, together with language labels from GlotLID, country and UN M49 region, MusicBrainz genre and style tags, prompt text, model identity and version, and links to the top reference tracks (Solak et al., 2 Oct 2025). Generated lyrics are classified with GlotLID, yielding 147 languages, of which 18 have at least 100 associated artists. Audio format and exact duration follow each platform’s defaults, and CLAP-compatible processing confirms that tracks are at least 10 seconds long.

This metadata design makes the benchmark suitable for multi-axis auditing. It supports aggregation by country, sub-region, language, model family, and genre class without requiring researchers to reconstruct provenance from external sources. That structure is central to the dataset’s role as an evaluation instrument rather than a collection of examples.

4. Evaluation formalism

GlobalDISCO evaluates generated music with three embedding models: PANNs, CLAP using the checkpoint music_audioset_epoch_15_esc_90.14, and MUQ-MuLan, which the paper describes as reflecting state-of-the-art tagging performance (Solak et al., 2 Oct 2025). For CLAP, embeddings are computed on 10-second windows with a 1-second hop, and per-track embeddings are obtained by mean aggregation over windows.

The principal alignment metrics are Frechet Audio Distance and Kernel Audio Distance. FAD compares generated and reference distributions, modeled as Gaussians in embedding space:

$$\mathrm{FAD} = ||\mu_r - \mu_g||_2^2 + \mathrm{Tr}(\Sigma_r + \Sigma_g - 2(\Sigma_r \Sigma_g)^{1/2}).$$

Lower values indicate better alignment. KAD is defined through Maximum Mean Discrepancy with a Gaussian RBF kernel:

$$\mathrm{MMD}^2 = E_{x,x'}[k(x,x')] - 2 E_{x,y}[k(x,y)] + E_{y,y'}[k(y,y')].$$

Again, lower is better.

To study mainstream drift, cosine similarity is computed between generated regional-genre tracks and mainstream-genre embeddings:

$s(x, y) = \frac{x \cdot y}{||x|| \, ||y||}.$

Distances are then aggregated across embedding models by ranking neighbors per model and summing ranks; the lowest aggregated rank identifies the closest mainstream neighbor. This procedure is used to determine whether outputs prompted as regional genres are closer to mainstream styles than to their intended reference distributions.

Normalization is performed per embedding/model combination via

$z = \frac{m - \mu}{\sigma},$

after which scores are averaged across embeddings and generation models to produce robust region- and genre-level comparisons. Region-level analyses average FAD and KAD across countries within UN M49 sub-regions, and heatmaps display mean z-scored scores. Genre-level analyses report mean normalized FAD and KAD across the ten most common mainstream genres versus selected regional genres. For auditing, the paper also defines absolute and relative disparity metrics:

$$\Delta = m_{HR} - m_{LR}, \qquad \delta = \frac{m_{HR} - m_{LR}}{m_{HR}}.$$

Formal significance testing, such as $t$-tests or bootstrapping, is not reported.

5. Empirical findings on bias

The central empirical result is the presence of marked regional disparities. Across PANNs, CLAP, and MUQ-MuLan embeddings, both FAD and KAD heatmaps show substantial performance gaps: generations for lower-resource regions, including Northern Africa, Sub-Saharan Africa, Southern Asia, and Western Asia, are markedly more out-of-distribution relative to their reference sets than generations for higher-resource regions such as Northern America (Solak et al., 2 Oct 2025). The paper emphasizes that these trends are consistent across metrics and embeddings.

Genre disparities follow a similar pattern. Mainstream genres achieve better alignment, reflected in lower FAD and KAD, than geographically niche or regional genres. Regional styles with substantial global online presence, such as southern hip hop and k-pop, perform near mainstream levels, whereas many regional and traditional styles, including soukous, pub rock, opera, and classical, score significantly worse. This suggests that geographic specificity alone is not the decisive variable; rather, the benchmark reveals an interaction between regional provenance and representation in widely available music corpora.

A further result is “mainstream drift” under regional prompting. For the six worst-scoring regional genres, Mureka and Suno often produce audio closer to pop or rock reference distributions than to the target genre’s references. The paper highlights opera and ghazal as especially clear examples: Suno’s outputs for these prompts are closer to pop than to their own genre references. The reported interpretation is that prompt-to-style mappings inside current systems are biased toward mainstream textures and forms.

The paper also reports convergence between embedding-based analyses and human perception. Qualitative examples made publicly available indicate that human listeners likewise perceive many regional-genre generations to be stylistically closer to mainstream neighbors than to their intended regional artist styles (Solak et al., 2 Oct 2025). The benchmark’s evidentiary claim is therefore not limited to one embedding family or one metric; it rests on cross-embedding consistency together with qualitative listening evidence.

6. Interpretation, access, limitations, and nomenclature

The paper interprets the observed disparities through several interacting factors. If model pretraining emphasizes high-resource regions and ubiquitous genres, learned priors will favor mainstream distributions. With 147 languages represented, GlobalDISCO further shows that lower-resourced languages and genres associated with underrepresented countries are harder to render faithfully, especially for traditional forms (Solak et al., 2 Oct 2025). The paper also argues that the agreement between PANNs, CLAP, and MUQ-MuLan strengthens the conclusion that the measured misalignment is not an artifact of any single embedding model.

GlobalDISCO is publicly available at https://huggingface.co/datasets/disco-eth/GlobalDISCO, with reference tracks supplied through links to LAION-DISCO-12M under Apache-2.0, and an accompanying website at https://a-b-solak.github.io/globaldisco/. The dataset includes generated audio per model, artist-level metadata, prompts and lyrics, and paired links to reference tracks. The paper describes the pipeline in detail, but scripts for data handling and evaluation are not specified. Ethical framing is explicit: the dataset is intended to reveal disparities without stigmatizing cultures, and researchers are urged to avoid reductive interpretations.

Several limitations are stated. Only four commercial systems are evaluated, all are black boxes, and generation hyperparameters are largely inaccessible. Reference selection depends on LAION-DISCO-12M and the top 10 most viewed tracks per artist, which may skew toward popular tracks and English-language metadata. Genre taxonomies derived from MusicBrainz plus tf-idf selection can conflate subgenres or miss local nuance, and Gemini-generated lyrics are only a proxy for authentic lyric-language and style coupling. Audio format, exact duration, and production settings also vary by platform.

A recurrent source of confusion is nomenclature. On arXiv, “DISCO” designates several unrelated systems, including distributed multi-domain SDN controllers (Phemius et al., 2013), an optical calibration instrument for neutrino detection (Rott et al., 2023), a web content discovery framework (Pham et al., 2019), and a sparse-communication distributed inference method (Qin et al., 2023). The 2025 music-bias benchmark is the source that explicitly introduces the name GlobalDISCO (Solak et al., 2 Oct 2025). In that sense, GlobalDISCO denotes not a generic “global DISCO” architecture, but a specific dataset and evaluation framework for measuring inequalities in AI-generated music at the level of regions, languages, countries, and genres.