Voxtar
Your voice changes before you notice. Voxtar measures the change.
A voice carries more than words. Some of what it carries shows up years early.
Something in the way a person speaks changes long before anyone notices it — the steadiness of a held vowel, the length of the pauses, how cleanly one syllable follows the next. The research literature has pointed at this for years. Almost nobody measures it.
Voxtar is the infrastructure for measuring it. Two minutes of recording, in the patient’s own language, from a phone at home or a handset in clinic. No appointment, no equipment, nothing to learn.
It is the part of my own training that bothered me most: the signal was there the whole time, and we had no way to listen to it.
Halleluyah Oludele · founderBuilt and demonstrated to consultants. No clinic is using it and no model is validated — there are no sensitivity, specificity or calibration figures anywhere in the library, so this page shows none. The next step is a supervised pilot.
Five short tasks — a held vowel, rapid syllables, a read passage, a cough, an s/z ratio. Two minutes, at home or in clinic.
Jitter, shimmer, pitch tremor, pause structure, articulation. Voxtar reads the sound, not the words.
A personal baseline, then drift away from it. The trend is the signal — a single reading means almost nothing.
What happens to a recording
Every capture, end to end
A recording that fails the quality gate — too noisy, clipped, too short — is thrown away rather than scored. A bad number is worse than a missing one.
How drift is flagged
SCHEMATIC — how the method works. NOT patient data; no patient data exists yet.
The baseline is the patient’s own trimmed average, with a floor on how small a change counts. A reading only matters once it leaves that band — and a short run of them matters more than any single day.
See the full numbers and sources
What pyzheimer measures
| Group | Measures |
|---|---|
| Acoustic features | Jitter, shimmer, maximum phonation time, s/z ratio, HNR, F0 statistics, CPP, formants and vowel space |
| Prosody | Pitch contour, rhythm, intonation, speech rate, pause structure |
| Voice quality | GRBAS-style bundles, dysphonia indices, glottal-source estimators |
| Signal quality | SNR estimation, clipping detection, segment-level gating |
| Baseline | Per-patient robust baseline — trimmed mean and MAD |
| Drift | Single-sample deviation plus short-window trend, with a per-feature floor |
Source: services/pyzheimer, read 19 September 2026. pyzheimer is a library, not a service — audio in, features out, no network in its core path, so a research notebook and the platform get identical numbers.
The five indication models
| Model | Reads for |
|---|---|
| Parkinson’s | Pitch micro-tremor, amplitude instability, vowel articulation boundaries |
| Cognitive decline | Prosody, pause structure, speech rate |
| Stroke-related dysarthria | Articulation and rate breakdown |
| Recurrent laryngeal nerve injury | Phonation and voice-quality collapse |
| Respiratory | Cough and breath-linked acoustics |
All five are wired into the analysis API. They are feature compilers built on published markers, not trained classifiers, and none carries a validation statistic yet.
The platform around it
| Piece | What it does |
|---|---|
| Patient app | Expo, guided daily capture with spoken instructions, phone-and-OTP sign-in |
| In-clinic session | Clinician-guided live conversation over LiveKit, biomarkers streaming to the console as it happens |
| Clinician console | Baseline, drift and per-cohort insight |
| Capture context | Ambient noise and SNR, time of day, device, weather and air quality |
| Record keeping | De-identified per patient, versioned per measurement |
Every feature row stores the exact pyzheimer version that produced it, so rolling the library forward never silently rewrites a patient’s history; baselines are recomputed per version rather than mixed across them.
What we do not claim
- No validation statistics. No sensitivity, specificity or calibration figure exists for any of the five models, so none is quoted.
- The models are not trained classifiers. They compile acoustic markers published in the literature; calling them diagnostic models would overstate what the code does.
- Nothing Voxtar emits is a diagnosis. Its outputs inform a clinician and never replace one, and there is no configuration in which that changes.
- No patient data. The drift figure above is a schematic of the method, not a real patient, and is labelled as such.
- A depression model appears in the library plan. It is not in the package and is not claimed here.
What a pilot has to settle, and what we will not say before it does.
- Does drift predict anythingWhether a flagged change precedes a clinical one. That is the whole question and it is unanswered.
- Against what standardEach model needs a validated comparator, per indication, on a consented cohort.
- Does it work hereRecorded on the phones people own, in a Nigerian clinic, in the language they actually speak.
- Voice is biometricConsent covers research use and voice-as-biometric separately, and recordings are de-identified.
- A clinician always decidesVoxtar hands over a measurement and a trend. It does not hand over a conclusion.
Capital to finish these and take them to market.
Six products across healthcare, education and AI. Where we have numbers, they are on the page. Where we do not, the page says so.