STraDa: Singer Traits Dataset

• STraDa is a large-scale dataset combining 30-second audio clips with rich lead-singer metadata for singing voice research.
• It comprises two subsets—automatic-strada for scalable model training and annotated-strada for controlled, balanced evaluation.
• The dataset enables both singer classification benchmarks and deep bias analysis using methods like CAV and TCAV.

STraDa, short for Singer Traits Dataset, is a large-scale public dataset for singing voice research that combines downloadable music audio files with rich lead singer metadata. It was introduced to address the limited availability of datasets that jointly provide public audio access, lead-singer-level annotations, and sufficient scale for both model development and bias analysis. STraDa is organized into two complementary subsets—automatic-strada and annotated-strada—and has subsequently been used not only for singer-focused benchmarking but also for diagnosing representational bias in pretrained music embedding models through Concept Activation Vectors (CAVs) and TCAV-style analyses (Kong et al., 2024, Gebhardt et al., 29 Sep 2025).

1. Scope, naming, and research role

In music and MIR research, STraDa denotes a dataset centered on lead singer traits rather than on generic track metadata. Its defining characteristic is the combination of downloadable 30-second excerpts with singer-level attributes such as gender, birth year, and active country, together with track-level attributes such as genre, language, and release date (Kong et al., 2024).

The dataset was designed to support at least two distinct research functions. First, it provides scale for supervised or fine-tuned singing-voice tasks, as demonstrated by the singer sex classification benchmark in the original paper. Second, it supports controlled subgroup analysis because its metadata enables evaluation across demographic and linguistic strata. This second function became especially salient in later work that used STraDa as the substrate for probing whether pretrained music representations entangle genre with non-musical concepts such as singer gender and language (Gebhardt et al., 29 Sep 2025).

A common misconception is that STraDa is uniformly balanced. The original dataset is not uniformly balanced across all attributes: automatic-strada is explicitly described as “not fully balanced,” whereas annotated-strada is deliberately balanced across selected variables for evaluation (Kong et al., 2024).

2. Dataset architecture and composition

STraDa comprises two subsets with different methodological roles. automatic-strada is the large-scale subset, while annotated-strada is the manually curated evaluation subset (Kong et al., 2024).

Subset	Composition	Access modality
automatic-strada	25,194 tracks; 5,264 unique singers; 169 hours total	Deezer API
annotated-strada	200 tracks from 200 unique singers; 1,200 three-second vocal excerpts	YouTube links with timestamps

The automatic-strada subset contains 25,194 tracks, each represented as a unique 30-second song excerpt, spanning 169 hours in total and covering 5,264 unique lead singers. It includes 25 genres and 35 languages. Track metadata comprises artist name, track name, album, release date, genre, language, and audio ID. Singer metadata comprises gender, birth year, active country, and unique singer ID. A distinctive inclusion criterion is that only tracks with a single, identified lead singer are retained, reducing annotation noise relative to datasets with uncertain vocalist attribution (Kong et al., 2024).

The annotated-strada subset contains 200 tracks from 200 unique singers and provides 1,200 three-second vocal excerpts, with six vocal-rich segments per track. It is manually curated and balanced across 2 genders (cis-male, cis-female), 5 languages (English, French, Mandarin, Spanish, German), and 4 age groups (20–34, 35–49, 50–64, 65+) (Kong et al., 2024).

The scale and distribution of automatic-strada are uneven but explicit. The five most frequent genres include pop (8,863 tracks), hip hop (5,277), rock (2,151), alternative (1,735), and electronic (1,160). The five main languages include English (11,422 tracks), French (4,555), Italian (1,978), Portuguese (1,570), and Japanese (1,429). Gender counts are 17,070 male-track instances, 7,933 female-track instances, and 191 non-binary-track instances, corresponding to 3,426 male singers, 1,827 female singers, and 11 non-binary singers (Kong et al., 2024).

3. Metadata acquisition and validation protocol

A central technical feature of STraDa is its cross-validated metadata pipeline. For automatic-strada, information is aggregated and reconciled across Deezer, MusicBrainz, Wikidata, and Discogs. Matching relies on alignment of album names, track names, and singer metadata across these sources, and inclusion is limited to cases where a single lead singer or “vocalist” role is clearly annotated (Kong et al., 2024).

The source-specific coverage is heterogeneous. Deezer provides artist, role, release date, gender, birth year, and active country; MusicBrainz provides artist and role, with partial release-date coverage; Wikidata provides artist and broad demographic fields, with partial role and release-date coverage; Discogs provides artist, role, and release date, with partial demographic coverage. Inconsistent or incomplete records are filtered out. For release-date harmonization, the chosen date is the earliest available across sources so as to approximate the actual recording date more closely. Birth year, gender, and active country are extracted via unique artist IDs and matched across MusicBrainz, Wikidata, and Discogs to resolve discrepancies (Kong et al., 2024).

The annotated-strada subset uses a different protocol. All 200 tracks are manually curated, annotated, and verified for gender, age, and language, and the six vocal-rich excerpts per track are selected specifically to maximize data diversity. Because direct audio sharing is restricted by copyright, access is provided through YouTube links and timestamps rather than through bundled audio files (Kong et al., 2024).

This dual design suggests a division of labor within the dataset: automatic-strada serves as the scalable source for model fitting and large-sample analysis, while annotated-strada serves as a controlled evaluation set with stronger balancing and manual verification.

4. Benchmarking in the original STraDa study

The original STraDa paper used the dataset to benchmark Singer Sex Classification (SSC) and to conduct subgroup bias analysis (Kong et al., 2024). For SSC, only cis-male and cis-female tracks were used, and non-binary instances were excluded because the task was formulated as a binary classification problem.

Four baseline systems were evaluated: F0 + kNN, MFCCs + MLP, a CNN trained on 3-second segments, and fine-tuned x-vector systems (X1 and X2). The F0 + kNN baseline used fundamental frequency histograms extracted with CREPE and a 5-nearest neighbor classifier. The MFCCs + MLP baseline used 13 MFCCs per song derived from separated voice tracks and a four-layer MLP. The CNN followed a prior singer-classification architecture and used both original and separated-voice inputs for augmentation. The x-vector systems were pre-trained speaker-recognition networks fine-tuned on singing voice, using 3-second Mel-spectrogram segments; X2 additionally alternated between polyphonic and voice-separated inputs via Spleeter-based augmentation (Kong et al., 2024).

Training used female singer segment upsampling to address class imbalance. Evaluation was performed on the 1,200 manually annotated segments in annotated-strada. The best-performing system was X2, with $89.8\%$ mean accuracy across 5 seeds (Kong et al., 2024).

Bias analysis in the original paper used recall at the subgroup level. The reported recalls for X2 were $85.6 \pm 2.5$ for female and $94.1 \pm 1.2$ for male singers; by age group, recalls were $91.3 \pm 1.5$ for 20–34, $93.6 \pm 1.0$ for 35–49, $86.0 \pm 1.0$ for 50–64, and $88.3 \pm 2.0$ for 65+; by language, recalls were $87.4 \pm 0.6$ for French, $88.8 \pm 2.5$ for English, $93.9 \pm 0.8$ for Mandarin, $85.6 \pm 2.5$0 for Spanish, and $85.6 \pm 2.5$1 for German (Kong et al., 2024). The paper identifies these results as evidence of gender bias, age bias, and language bias in the benchmarked system.

5. Use in concept-based bias diagnosis of music embeddings

A later study, “Beyond Genre: Diagnosing Bias in Music Embeddings Using Concept Activation Vectors,” used STraDa as the foundational dataset for probing how non-musical singer attributes influence genre representations in pretrained models (Gebhardt et al., 29 Sep 2025). In that work, the automatic-strada subset was used with metadata for over 25,000 tracks, and 30-second audio previews for 22,168 tracks were collected via the Deezer API. The study further augmented the data with 251 tracks from Deezer playlists targeting underrepresented genre–gender combinations, with manual checking of genre and singer attributes (Gebhardt et al., 29 Sep 2025).

The experimental design centered on binary concept probes for gender (male/female) and for the seven most common languages: English, French, Italian, Portuguese, Japanese, Spanish, and German. To reduce confounding between concept labels and genre, the authors used stratified sampling such that, within each genre, equal numbers of positive and negative samples were drawn for a given concept. When data was sufficient, training used up to 50 samples per (label, genre) cell, with the remainder reserved for testing; scarce cells were downscaled proportionally. Train and test sets were disjoint, and dominant subgroups were capped so that cells such as “male rock” would not dominate the learned direction (Gebhardt et al., 29 Sep 2025).

For CAV learning, the study extracted final-layer embeddings from STraDa tracks and trained a linear classifier (logistic regression) of the form

$85.6 \pm 2.5$2

where $85.6 \pm 2.5$3 is the embedding, $85.6 \pm 2.5$4 is the concept vector, and $85.6 \pm 2.5$5 is the bias term. CAVs with insufficient accuracy were discarded on the grounds that the concept was not linearly encoded for that model (Gebhardt et al., 29 Sep 2025).

The dataset supported evaluation of four state-of-the-art models: MERT, Whisper, MuQ, and MuQ-MuLan. For each concept, the study trained 500 CAVs on independently sampled balanced training subsets comprising 25% of the data, then computed TCAV scores by genre. Significance was assessed using a two-sided t-test against the null value 0.5, with Bonferroni correction for multiple comparisons (Gebhardt et al., 29 Sep 2025).

Using these STraDa-based controls, the study reported that all four embedding models displayed statistically significant genre-specific biases with respect to both gender and language, despite genre-balanced splits. The reported patterns were model-specific: MERT showed strong male dominance for genres such as Metal and Hip-Hop and positive female association in Electronic and Soul; Whisper showed subtler and divergent genre patterns; MuQ and MuQ-MuLan showed similar directional trends, with MuQ-MuLan exhibiting stronger language/genre entanglement under joint text supervision. A language-focused case study found that Portuguese CAVs aligned strongly with genres such as Latin American Music in MERT and MuQ(-MuLan), while Whisper displayed less entanglement (Gebhardt et al., 29 Sep 2025).

The same study also used STraDa to test post-hoc debiasing by manipulating concept vectors. In a reported example, a Hip-Hop direction was adjusted by interpolating with a female vocals CAV, and this directly reduced male overrepresentation at the top of a gender-balanced test set, with nearly linear mitigation (Gebhardt et al., 29 Sep 2025).

6. Limitations, access, and significance for MIR

STraDa’s limitations are explicit. Non-binary representation in automatic-strada is small—11 singers and 191 tracks—and the original paper notes that the accuracy of non-binary annotation is uncertain because source databases document this inconsistently. The dataset also exhibits cultural skew, including overrepresentation of Western pop genres, and annotated-strada includes only cis-gender singers (Kong et al., 2024). A later study therefore restricted its gender probing to binary gender because of data sparsity (Gebhardt et al., 29 Sep 2025).

These limitations matter methodologically. They imply that STraDa is especially strong for controlled analyses of major subgroups represented in current music metadata ecosystems, but less suitable for fine-grained fairness claims about marginalized groups whose documentation is sparse. A plausible implication is that future expansions with stronger representation of non-binary and other underrepresented populations would materially improve the dataset’s value for bias analysis.

Despite these constraints, STraDa occupies a distinctive place in MIR because it combines publicly accessible audio, lead-singer disambiguation, cross-validated metadata, and a balanced manually curated subset within a single resource. The original paper positions it for Singer Sex Classification, Singer Recognition and Identification, Singer Age Detection, and Bias Analysis across gender, age, language, while later work shows that the same metadata richness supports post-training diagnosis and mitigation of representational bias in general-purpose music embeddings (Kong et al., 2024, Gebhardt et al., 29 Sep 2025).

STraDa is publicly available through Zenodo at https://zenodo.org/records/10057434 (Kong et al., 2024).