VoxClinBench (Hugging Face dataset mirror)

Cross-lingual, cross-disease clinical voice biomarker benchmark. This Hugging Face dataset mirror ships the split manifests, Croissant metadata, datasheet, and reference baseline prediction CSVs.

Raw audio is NOT distributed here. Each of the five upstream corpora must be obtained directly from its provider under that corpus's data use agreement (DUA). See the GitHub mirror for the evaluation harness and the voxbench fetch CLI.

v0.4.1 release snapshot (2026-04-16)

• 24 task registrations across 5 corpora (Bridge2AI-Voice v3.0, NeuroVoz, SVD, E-DAIC, MODMA) in 4 languages (en, es, de, zh)
• 23 scored rows (Family A physical n=18, Family B psychiatric n=5)
• 1 zero-shot cross-lingual reference row (A24 MODMA Mandarin depression; outside the Holm family)

What's new in v0.4.1

• Task ID A26 renumbered to A24 for the MODMA cross-lingual reference row (old A24/A25 regression tasks retired from the scored registry; the per-seed Pearson r results are preserved in the paper's supplementary "Exploratory Regression Analyses" section without task IDs).
• A9 neurovoz.parkinsons: 3-seed headline (0.955 ± 0.041) supplemented to 5 seeds: 0.936 ± 0.039.
• A10 svd.vf_paralysis: 3-seed headline (0.929 ± 0.021) supplemented to 5 seeds: 0.936 ± 0.019.
• A24 modma.depression: legacy single-head mean 0.799 superseded by 5-seed WavLM-L9 zero-shot: 0.680 ± 0.092.
• macro_physical_18 recomputed with updated A9/A10 values (artifacts/eda/dual_family_macro.json).
• §Methods backbone policy: reported baseline is now "reasonable-effort per-task with whichever SSL backbone we evaluated with sufficient seeds"; per-task evaluated-backbone matrix added as supplementary \ref{ssec:backbones-matrix}.

Task roster

#	Task ID	Family	Kind	Corpus	Lang	n_pos	n_neg
A1	b2ai.parkinsons	A	within	bridge2ai	en	73	689
A2	b2ai.airway_stenosis	A	within	bridge2ai	en	55	707
A3	b2ai.vf_paralysis	A	within	bridge2ai	en	60	702
A4	b2ai.laryngeal_dystonia	A	within	bridge2ai	en	50	712
A5	b2ai.cognitive_impair	A	within	bridge2ai	en	45	717
A6	b2ai.benign_lesions	A	within	bridge2ai	en	58	704
A7	b2ai.mtd	A	within	bridge2ai	en	52	710
A8	b2ai.chronic_cough	A	within	bridge2ai	en	41	721
A9	neurovoz.parkinsons	A	external	neurovoz	es	53	54
A10	svd.vf_paralysis	A	external	svd	de	213	300
A11	svd.hyperfunctional_dysphonia	A	external	svd	de	199	300
A12	svd.laryngitis	A	external	svd	de	128	300
A13	svd.functional_dysphonia	A	external	svd	de	108	300
A14	svd.psychogenic_dysphonia	A	external	svd	de	80	300
A15	svd.contact_pachydermia	A	external	svd	de	63	300
A16	svd.reinke_edema	A	external	svd	de	54	300
A17	svd.dysodia	A	external	svd	de	54	300
A18	svd.vf_polyp	A	external	svd	de	40	300
A19	b2ai.depression	B	within	bridge2ai	en	44	718
A20	b2ai.ptsd	B	within	bridge2ai	en	38	724
A21	b2ai.adhd	B	within	bridge2ai	en	47	715
A22	edaic.depression	B	external	edaic	en	82	193
A23	edaic.ptsd	B	external	edaic	en	87	188
A24	modma.depression	B	zero-shot reference	modma	zh	23	29

A23 PTSD binary labels derive from the PTSD_label column (cutoff-based) in E-DAIC's detailed_lables.csv.

Upstream fetch

Use the voxbench GitHub CLI:

pip install voxbench
voxbench fetch bridge2ai   # PhysioNet credentialed (hard login wall)
voxbench fetch edaic       # USC/ICT EULA (AVEC'19; subsumes DAIC-WOZ)
voxbench fetch svd         # CC-BY-4.0 Zenodo mirror (record 16874898)
voxbench fetch neurovoz    # CC-BY-NC-ND-4.0 Zenodo access (record 10777657)
voxbench fetch modma       # CC-BY-NC-4.0 Lanzhou University form

NeuroVoz split regeneration (CC-BY-NC-ND-4.0 ND clause)

NeuroVoz's CC-BY-NC-ND-4.0 license forbids redistributing derivatives. A subject-ID split manifest is a derivative. VoxClinBench therefore does not redistribute the NeuroVoz split as a CSV/JSON. Regenerate locally from your fetched copy:

python -m voxbench.splits.neurovoz_splitter \
    --data_dir /path/to/neurovoz \
    --seed 42 \
    --out_dir ./splits/neurovoz.parkinsons/

The script is deterministic: given the same NeuroVoz filesystem snapshot and the same seed, it produces the same split manifest VoxClinBench-Base used for the A9 leaderboard row, without copying any NeuroVoz content into the benchmark repo.

Protocol

• Subject-disjoint stratified 80/20 train/test; internal 80/20 train/val
• Five split seeds: {42, 1, 2, 3, 4}
• Primary metric: AUROC
• Secondary: AUPRC for rows with positive prevalence <15%
• Bootstrap: 2000 resamples for 95% CIs
• Between-submission test: DeLong on per-subject probabilities
• Correction: Holm-Bonferroni at alpha=0.05 within each family

Retraction ledger (what used to be here and is not)

• old A12 b2ai.psychiatric_history — excluded v0.3.3 (composite self-report whose positives are a union of A19-A21 positives in the same cohort; not an independent biomarker).
• old A15 daicwoz.depression — retracted v0.3.1 (189/189 subset of E-DAIC; not independent).
• LODO rows (cross-lingual B1-B6) — removed v0.3.2 (training-corpus choice is submission-disclosed, not benchmark-prescribed).
• SVD leukoplakia and SVD presbyphonia — scoped out v0.4.0 (n_pos < 40 threshold; presbyphonia also has age-HC confound).
• old A24 edaic.phq8_regression and old A25 edaic.ptsd_severity — retired v0.4.1 from the scored registry; exploratory Pearson r results preserved in the paper's supplementary "Exploratory Regression Analyses" section without task IDs. Task ID A24 reassigned in v0.4.1 to the MODMA cross-lingual reference row (previously A26).

See CHANGELOG §0.4.1 for the full release delta.

Citation

@article{voxclinbench2026,
  title={VoxClinBench: A Clinical-Voice Benchmark with Cross-Lingual Expansion},
  author={Anonymous},
  year={2026},
  note={Submitted to NeurIPS 2026 Datasets \& Benchmarks Track},
  url={https://github.com/voice-bench-submission/voxclinbench},
}

License

• Code: MIT
• Splits (where shipped), docs, Croissant: CC-BY-4.0
• Upstream corpora: individually licensed; see fetch table above
• NeuroVoz A9 manifest: regenerated locally via splitter script (CC-BY-NC-ND-4.0 ND compliance)