How should sensitivity and specificity be interpreted?

Sensitivity is the proportion of participants with the reference condition whom the model correctly identifies; specificity is the proportion without the condition whom it correctly identifies. These values apply to the evaluated cohorts, reference measures and recording conditions shown here and should not be assumed to transfer unchanged to every population, device or intended use.

Can results be compared directly across conditions?

No. Each condition is evaluated against its own reference measure, cohort, decision threshold and validation design. Sensitivity and specificity therefore describe performance within that specific evaluation; they should not be used to rank conditions or infer that one condition is easier to identify than another.

How does language affect voice biomarker performance?

Virtuosis models have been trained using recordings from many diverse languages, so we expect them to achieve similar performance in languages not listed in the table. We nevertheless monitor performance when deploying in a new language. Our validation datasets also include additional languages represented by smaller groups, typically 10–30 participants. To avoid overinterpreting these smaller samples, the table lists only the main languages evaluated for each condition.

Voice biomarker guide

Clinical validation

How Virtuosis validates voice biomarkers

Four principles guide every validation:

Reference standard

We compare model outputs with a clinically meaningful endpoint or validated assessment appropriate to the intended use.

Representative data

We evaluate data reflecting the intended population, languages, devices and realistic recording conditions.

Independent evaluation

We test on held-out, external or prospective data that did not influence model development.

Transparent and robust reporting

We document sensitivity, specificity, AUROC or calibration as appropriate, together with uncertainty, subgroup, language, device and noise checks, and known limitations.

Virtuosis clinical validation results

The table reports sensitivity and specificity for eight intended use cases, alongside their reference measures, cohort sizes and evaluated languages. Cohorts were balanced by class (healthy control versus pathological), age and sex, plus weight and smoking habits where applicable.

Condition	Reference measure	Sensitivity	Specificity	Participants	Languages
ConditionStress	Reference measureSalivary cortisol	Sensitivity98%	Specificity92%	Participants400	LanguagesFrench, Italian, English, Spanish, German, Portuguese and Russian
ConditionAnxiety	Reference measureGAD-7	Sensitivity61%	Specificity79%	Participants1,132	LanguagesFrench, Italian, English, Spanish, German, Portuguese, Chinese and Russian
ConditionDepression	Reference measurePHQ-9	Sensitivity77%	Specificity83%	Participants1,933	LanguagesFrench, Italian, Chinese, English, Spanish, German, Portuguese, Malay and Russian
ConditionParkinson’s disease	Reference measureUPDRS	Sensitivity98%	Specificity96%	Participants1,621	LanguagesFrench, Slovak, English, Italian, Spanish, German, Arabic and Turkish
ConditionAlzheimer’s disease	Reference measureMMSE	Sensitivity97%	Specificity92%	Participants2,488	LanguagesEnglish, Greek, French, Slovak, Chinese, German and Italian
ConditionMild cognitive impairment	Reference measureMoCA	Sensitivity81%	Specificity76%	Participants2,566	LanguagesSlovak, Mandarin Chinese, English, German and Italian
ConditionRespiratory function	Reference measureCOPD, COVID-19 or asthma diagnosis	Sensitivity72%	Specificity82%	Participants30,312	LanguagesSpanish, English and French
ConditionType 2 diabetes	Reference measureHbA1c levels	Sensitivity73%	Specificity70%	Participants1,440	LanguagesSpanish, English and French

Supported voice biomarkers can be integrated through the Virtuosis API.

Conference abstracts and posters

Gervaise L et al. Voice biomarkers for multi-condition health screening using a short natural speech sample: a real-world validation. International Vocal Biomarkers Conference (IVBC). 2026.

Gervaise L et al. Speech-based biomarkers for scalable detection of Alzheimer’s disease, mild cognitive impairment, and Parkinson’s disease. European Journal of Neurology. 2026.

De Kalbermatten C, Gervaise L, Qorraj M. Family voice in rare diseases (FAVOUR-RD). European Conference on Rare Diseases (ECRD). 2026.

Gervaise L. Evaluation of voice-based screening for depression across controlled and real-world speech datasets. Voice AI Symposium. Frontiers. 2026.

Mekki M, Gervaise L. Voice-based depression detection: a bio-inspired and language-invariant model. Applied Machine Learning Days (AMLD). 2026.

FAQ

Planning a project?

Request a demo