VoxClinBench

Anonymous submission for NeurIPS 2026 Evaluations & Datasets Track.

First cross-lingual, cross-disease clinical voice biomarker benchmark. Six corpora (Bridge2AI-Voice v3.0, NeuroVoz, SVD, DAIC-WOZ, E-DAIC, MODMA), four languages (en, es, de, zh), 22 ranked subject-disjoint classification tasks + 1 PHQ-8 regression scoping entry.

What this repo contains

• splits/ — 6 partial subject-ID manifests (test subjects; train/val regenerable via voxbench.data.make_splits after upstream fetch).
• predictions/ — 66 per-seed probability CSVs (11 external-corpus seed-0 baselines + 55 B2AI Tier-2 5-seed × 11-disease canonical CSVs regenerated from our saved checkpoints). Each CSV: exactly two columns subject_id,predicted_prob (labels filtered out per DUA).
• croissant.json — NeurIPS 2026 E&D Croissant metadata.
• README.md, license.

Raw audio is not redistributed — use the voxbench fetch CLI from the companion GitHub mirror to get per-corpus URLs + credential check.

Companion GitHub repo (harness, training, baselines)

https://github.com/voice-bench-submission/voxclinbench

Two workflows supported:

# A. Submit your own model (pure numpy+scipy+sklearn, no torch)
pip install voxbench
voxbench eval --task <id> --predictions your.csv --labels upstream.csv

# B. Reproduce / retrain VoxClinBench-Base (heavy deps)
pip install "voxbench[train]"
python -m voxbench.train --help

Fetch corpora

Corpus	License	Access
bridge2ai	PhysioNet credentialed	hard login wall
daicwoz	USC/ICT EULA	HTTP public; EULA on download
edaic	USC/ICT EULA (AVEC'19)	HTTP public; EULA on download
svd	CC BY 4.0 (Zenodo mirror)	publicly downloadable
neurovoz	CC BY-NC-ND 4.0	Zenodo access request
modma	CC BY-NC 4.0	Lanzhou University form

License

MIT (code) / CC BY 4.0 (manifests, docs). Each upstream corpus retains its native license.