Every corpus the field uses to claim a disease can be heard in a voice — organised by what the sound is supposed to reveal, and scored on the one question the published surveys do not ask: could a result on this dataset mean anything?
The claim column. A survey tells you a dataset exists. This column tells you what a number measured on it would be worth, under a rigor contract that requires speaker-disjoint splits and a confound baseline:
These are this program's judgements about use here, not criticisms of the corpora or their authors — several of the BLOCKED rows are the best-designed datasets in the field, and the SAP row is the model everyone else should follow.
The sound-producing organ itself is altered -- paralysis, polyp, oedema, tumour, surgical scarring. The shortest causal chain in the field and where the evidence is strongest.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| SVD (Saarbrücken Voice Database) 2007–on this host — decoded: 28,509 recordings / 1,679 speakers | 71 voice pathologies vs healthy | 2,043 speakers (1,356 pathological / 687 healthy) | German | clinically validated | open free | UAR 85.22 combined — arXiv:2410.10537 | EVALUATION THE PRIMARY TARGET. Real speaker ids, 1,679 decoded here. But only 12 of the 71 pathologies have ≥30 speakers and 19 have exactly one — so the powered task is binary, not 71-way. |
| Bridge2AI-Voice v3.1.0 2026 2026-05 | 5 cohorts incl. voice disorders, neuro, mood, respiratory, pediatric | 833 participants (v3.1); VoiceFM reports 984 / 40,056 recordings / 176 h | English (North America) | clinician-confirmed + validated questionnaires | DUA credentialed + signed DUA; raw audio needs a separate DACO application | VoiceFM contrastive baseline (medRxiv 2026.05.28); laryngeal-lesion model, Front. Digit. Health 2026 | BLOCKED The most ambitious purpose-built corpus in the field. PhysioNet tier ships DERIVED FEATURES ONLY (spectrograms/MFCC/PPG) — raw audio is a second committee. Paperwork-bound, not compute-bound. |
| RRP-Voice 2026 2026-06 | recurrent respiratory papillomatosis | longitudinal, per-subject follow-up up to 10 years | English | synchronous laryngoscopic ground truth at each visit | on request see paper | benchmark released with the paper | SCREENING Unusually strong labels — an endoscopic look at the larynx at every recording, which almost no corpus has. But longitudinal repeats mean subject leakage is near-guaranteed under naive splits. |
| VOICED 2018 | voice pathology | 208 voices (150 pathological / 58 healthy) | Italian | clinical | open open | — | SCREENING Small and single-vowel, but open and in a different language from SVD — which makes it the natural cross-corpus destination for an SVD-trained model. |
| PVQD (Perceptual Voice Qualities Database) 2020 | dysphonia severity (CAPE-V) | 296 recordings | English | expert perceptual ratings | open open | — | SCREENING Severity ratings rather than diagnosis, and studio conditions make results optimistic. Useful as a regression target, not a detection benchmark. |
Speech needs fine neuromuscular control, so hypokinetic dysarthria appears early in PD. A longer causal chain than the larynx family but a well-replicated one.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| PC-GITA 2014 | Parkinson's disease | 100 speakers (50 PD / 50 HC), sex-balanced | Spanish (Colombian) | neurologist diagnosis + MDS-UPDRS, Hoehn & Yahr | on request on request from authors | early-PD F1 0.73 / AUC 0.80 (DDK) — arXiv:2605.14066 | SCREENING n=100 is far below this program's floor. Its value is the metadata: PC-GITA and NeuroVoz are the only two corpora carrying enough staging detail to isolate EARLY-stage PD. |
| NeuroVoz 2024 | Parkinson's disease | 112 speakers (54 PD / 58 HC), 2,977 audio files | Spanish (Castilian) | clinician diagnosis, UPDRS | DUA Zenodo restricted + DUA | shares the EarlyPD benchmark with PC-GITA | SCREENING ON-medication state only, so it cannot separate disease from treatment response. Pairs with PC-GITA for a Spanish cross-corpus test. |
| mPower 2016– | Parkinson's disease | tens of thousands of voice tasks, longitudinal smartphone cohort | English | SELF-REPORTED PD status | on request Synapse qualified-researcher terms | — | SCREENING By far the largest PD voice cohort, and the most dangerous: self-report labels plus heavy repeated measures per participant. Reported accuracies here have historically collapsed once splits were made subject-disjoint. |
| MDVR-KCL 2019 | Parkinson's disease | small (mobile-device recordings, PD + controls) | English | clinical | open open (Zenodo) | — | SCREENING Mobile-device capture makes it a realistic acquisition-shift partner for studio-recorded PD corpora. |
Language production degrades before motor speech does, so the signal is as much linguistic (word-finding, semantic density) as acoustic. This is why text models beat audio models here — and why the shortcut risk is highest.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| PROCESS-2 2026 2026-05 | HC / MCI / dementia 3-way + MMSE regression | 400 participants (200/150/50), 1,200 recordings, ~21 h | British English | NHS memory-service diagnosis; MMSE for only 174/400 | gated HuggingFace gated repo + DUA | macro-F1 0.59 (3-way), F1 0.85 (2-way), MMSE RMSE 3.87 | BLOCKED The highest-headroom cognitive task available (macro-F1 0.59 leaves real room). But 400 participants fails the ≥500/class bar and MMSE is missing for 56.5%. |
| ADReSS / ADReSSo / ADReSS-M / TAUKADIAL 2020–2024 | Alzheimer's dementia, MCI, MMSE regression | ~156 subjects (ADReSS), balanced for age and sex BY DESIGN | English; ADReSS-M multilingual | clinical diagnosis + MMSE | gated DementiaBank consortium password + T&C | F-score ~0.79 (with age+sex as inputs); ADReSSo acoustic ~0.76 | BLOCKED The one family that already age- and sex-matches by design — the correct response to the confound this program studies. Note the headline uses age+sex AS INPUTS, so it is not a confound-free number either. Balancing removes age/sex; session, administration and transcription effects remain. |
| PARLO Dementia Corpus 2026 2026-03 | Alzheimer's disease | German multi-centre | German | clinical | on request see paper | [UNVERIFIED] | BLOCKED Non-English dementia corpora are rare; multi-centre design means site effects must be measured, not assumed away. |
Direct motor impairment of articulation from PD, ALS, cerebral palsy, stroke. Severity, not diagnosis, is usually the target.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| SAP (Speech Accessibility Project) 2024–2025 | dysarthria: Parkinson's, ALS, cerebral palsy, stroke, Down syndrome | 524 participants, ~415 h | English | etiology-labelled + severity | DUA UIUC data agreement | Interspeech 2025 SAP Challenge (ASR WER) | BLOCKED THE POSITIVE EXAMPLE of the whole survey: explicitly designed for speaker-INDEPENDENT evaluation, at a scale nothing else in this table reaches. If more corpora were built this way, this program would have less to audit. |
| TORGO / UASpeech 2008–2012 | dysarthria | TORGO 15 speakers; UASpeech 19 speakers × 765 words | English | severity ratings | open open / registration | — | NEVER 15 and 19 speakers with thousands of utterances each. Any random split leaks speaker identity almost surely; the reported accuracies are largely speaker-recognition. Not a generalisation benchmark at any n. |
Airway obstruction and inflammation change cough and breath acoustics. Also the field's canonical cautionary tale — see the UK ONS row.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| Coswara 2020–2022on this host — decoded: 610 files / 72 participants | COVID-19 | 2,635 individuals (1,819 neg / 674 pos / 142 recovered) | multiple (India) | MOSTLY SELF-REPORTED | open open | AUC ~0.92 intra-dataset | SCREENING Real participant ids, so speaker-disjoint splits are possible — but the returning-user flag is missing for 680 of 2,746, making it submission-disjoint rather than person-disjoint. Intra-dataset AUC ~0.92 collapses cross-corpus. |
| COUGHVID 2021on this host — decoded: 13,535 clips (10,132 healthy / 2,683 symptomatic / 720 COVID-19) | COVID-19 | >25,000 crowdsourced cough recordings | global | self-report + expert-labelled subset | open open (CC via Zenodo) | AUC ~0.93 intra-dataset | NEVER PERMANENTLY DEMOTED HERE. 13,535 decoded rows, 13,535 unique uuids, ZERO participant identifiers. A 'speaker-disjoint' split over recording uuids is not speaker-disjoint, and there is no way to detect the violation. Pretraining and OOD probing only. |
| UK ONS / Turing vocal audio 2024 | COVID-19 | 67,842 individuals, 23,514 PCR-positive | English (UK) | PCR-REFERENCED — the strongest labels in the family | gated restricted | AUC 0.846 unadjusted → 0.619 after matching on confounders | BLOCKED THE MOST IMPORTANT ROW IN THIS TABLE. The largest, best-labelled respiratory corpus ever built, and when its authors matched on recruitment confounders the headline fell from 0.846 to 0.619. That single number is the empirical case for this entire program. |
| ICBHI 2017 respiratory sound 2017 | crackles / wheezes (acoustic events, not diagnosis) | 126 subjects / 920 recordings | — | expert annotation | open open | ICBHI score (event-level) | SCREENING Event-level, not nosological — it detects a sound, not a disease. Fixed equipment per site makes device a near-perfect site proxy. |
Prosody, pause structure and speaking rate shift with mood state. The weakest causal chain in the field, and the one where interview-protocol artifacts do the most damage.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| DAIC-WOZ / E-DAIC (AVEC) 2014–2019 | depression, PTSD, anxiety | 189 sessions | English | PHQ-8 | gated USC EULA | — | NEVER The interviewer-prompt shortcut: the questions asked of depressed participants differ, so a model can score well by classifying the INTERVIEWER. Audited again in 2026 (arXiv:2605.23977) with worse news — see below. |
| CMDC / ANDROIDS / MODMA / PDCH 2026 2020–2023 | depression | small clinical-interview corpora | Chinese / Italian / Chinese / — | clinical + questionnaire | on request varies, mostly on request | audited in arXiv:2605.23977 | NEVER The 2026 multi-probe audit found models strong on CMDC and ANDROIDS fail to transfer zero-shot, and that text beats audio consistently across all five seeds — i.e. these benchmarks are measuring interview content, not voice. |
| Voice Biomarkers for Depression & Anxiety 2026 2026-05 | depression, anxiety | ~65,000 utterances / 23,000+ subjects (train); ~5,000 subjects (eval) | English (US-representative) | clinical instruments | not released DATA PROPRIETARY; model released on HuggingFace | 71% sensitivity and specificity | NEVER By far the largest mental-health voice cohort ever reported — and the data is not released, so the 71% cannot be audited by anyone. The released model is usable; the claim is not checkable. |
Rather than one disease, these ask what a single voice sample carries across many conditions at once — the honest version of the 'voice as a vital sign' claim.
| dataset | conditions | size | language | labels | access | published result | claim & why |
|---|---|---|---|---|---|---|---|
| SpeechDx 2026 2026-06 | 27 tasks across 12 datasets, organised by speech-production stage (conceptualization / formulation / articulation) | aggregates 12 corpora | multiple | inherited from constituent corpora | on request see paper | large-scale speech models are the strongest overall baselines | BLOCKED THE MOST IMPORTANT NEW RELEASE for this program. Its headline conclusion is that NO current representation generalises reliably across the clinical speech landscape — independent, and from a different direction, of what F3 found here. Its production-stage taxonomy is also the right way to organise this table. |
| HPP-Voice 2025-05 | 15 phenotypes across 6 systems: sleep apnea, insomnia, asthma, smoking, sinusitis, anemia, thyroid, migraine, depression, anxiety, COVID, allergy … | 7,188 recordings from 6,760 adults, mean age ~52 | Hebrew | linked deep clinical phenotyping | on request qualified researchers on request | sleep apnea (males) AUC 0.64 ± 0.03 vs MFCC 0.56 vs DEMOGRAPHIC BASELINE 0.57 | BLOCKED THE REALISM CHECK. The only large study in this table that reports a demographic baseline next to its headline — and once it does, the best effect is 0.64 against a 0.57 bar. That is what an honest voice-phenotype effect size looks like. |
| RespiraMFM 2026 2026-06 | 5 major respiratory diseases | 7 real-world datasets | multiple | inherited | on request see paper (ACL 2026 Main) | +9.15% AUROC supervised, +20.98% zero-shot | BLOCKED Multimodal by design — it fuses audio with patient history and symptoms. Worth flagging that adding symptoms to a respiratory model is exactly the confound the UK ONS study isolated; the gain must be attributed carefully. |
| Voice EHR 2024 | multimodal audio + EHR | — | multiple | EHR-linked | on request see paper | — | BLOCKED Represents the collection-protocol direction: build the corpus around the health record rather than retrofitting labels onto found audio. |
| study | id | why it matters here |
|---|---|---|
| SpeechDx: A Multi-Task Benchmark for Clinical Speech AI Bhalla, Kieu, Merchant, de Lara, Mariakakis · 2026-06-15 | arXiv:2606.17339 | 27 tasks over 12 datasets. Concludes NO current representation generalises reliably across clinical speech — converging with this program's F3 from a completely different direction. |
| A Multi-Probe Audit of Clinical-Interview Depression Detection Benchmarks Ishikawa & Duke · 2026-05-13 | arXiv:2605.23977 | The closest published work to this program's method. Audits DAIC/E-DAIC, CMDC, ANDROIDS, MODMA, PDCH and finds the official E-DAIC leaderboard is essentially noise: the best CV configuration ranks 20th on the official test, the test winner ranks 41st by CV, and top-3 overlap is ZERO. Text beats audio on symptom-dense segments across all seeds. |
| Systematic FAIRness Assessment of Open Voice Biomarker Datasets Mahapatra & Mahapatra · 2025-08-14 (TSD 2025) | arXiv:2508.14089 | Scores 27 public voice-biomarker datasets on FAIR maturity. Findability is uniformly high; accessibility, interoperability and reusability vary wildly. Explains why half the rows above say BLOCKED. |
| Vocal audio for COVID-19 screening (UK ONS / Turing) Turing-RSS Health Data Lab · 2024 | Sci Data s41597-024-03492-w | AUC 0.846 unadjusted → 0.619 after matching on recruitment confounders, on 67,842 PCR-referenced individuals. The single strongest empirical justification for measuring the margin above a confound baseline. |