A senior-focused speech corpus for voice health AI
A multilingual speech dataset dedicated to older adults, recorded under harmonised protocols and ethical oversight, built for teams training voice models toward real-world clinical and well-being use.
Seven languages collected against one protocol, so cross-language modelling is meaningful from the start.
Built to be balanced before you start modelling
The corpus was recruited against a specification, which is what makes the cross-language comparisons hold.
50 to 80+
Speakers spanning four decades of older adulthood, with consistent representation across each age band.
Balanced cohorts
Speaker recruitment is balanced across self-reported gender categories within each language.
Consistent ratios
Speaker ratios are held consistent across languages, which reduces confounding by cohort composition in cross-language comparisons.
Recording standards
Consistent acoustic environments and recording standards, with a quality check on every recording.
Seven languages under one protocol
Each language uses locally appropriate stimuli and licensed materials where required, with metadata aligned across the full corpus.
Six elicitation types per speaker
Fixed stimuli are what make speech comparable across speakers, sites and languages; without them, two recordings differ both because the speakers differ and because the elicitation did.
Picture description
Connected speech elicited from a visual stimulus, supporting analysis of fluency, content and discourse structure.
Narrative recall
Short structured narratives capturing memory and language organization.
Verbal fluency
Semantic and phonemic generation tasks.
Read passages
Language-matched standardised passages for prosodic and acoustic analysis.
Phonation and paralinguistic tasks
Sustained phonation and elicitations capturing voice quality, pitch and articulation.
Spontaneous speech
Open-ended prompts for naturalistic speech.
A clinical anchor, without identifiable medical records
Each speaker is documented with structured contextual metadata: a clinical anchor for every speaker, with no identifiable clinical records needed.
Screening scores
Captured through widely recognised screening instruments, selected for cross-language comparability.
Self-reported health
Self-reported indicators relevant to respiratory, cardiovascular and pain status.
Mood and well-being
Standardised mood and affective screening scores.
Demographics
Age, self-reported gender, language background and education.
Recording context
Device class, acoustic environment, task code and consent version: the acquisition-chain variables that move model outputs.
Beyond cognition
Physical and affective context alongside cognitive status.
Consent that covers training and commercial use
Speech with health metadata is special-category data under GDPR. The usual constraint is consent scope: whether the people who recorded it agreed to the intended use.
Research, licensing and model training
Consent is layered to cover research use, commercial licensing and model training, with subject rights honoured throughout the data lifecycle.
Independent review
The study protocol goes through independent ethics review at every site.
Provenance pack
DPIA, ROPA, DPA and the full compliance documentation are shared with licensees at onboarding, to support your own data-protection review.
Identity verification
Speaker identity is verified, ensuring data integrity and preventing duplication across sources.
Tell us what you're building
The languages you need, the population, and what the model has to do downstream.
Discuss a licence →