Research previewVoice biomarkers, measured from ordinary conversation
Your voice is a vital sign.
Speech carries measurable signals of how you are doing — how steadily you hold a pitch, how you pause, how clearly you articulate. Harmony measures them from a thirty-second conversation and tracks how they move over time.
Harmony is a wellness tool, not a diagnostic device. It reports what it measures and nothing more.
Session reading
0:31
- Pitch stability
1.2%
jitter
- Harmonic clarity
12.4
dB
- Speech rate
148
words / min
- Mean pause
480
ms
How it works
Thirty seconds of speech, measured the same way every time.
01
Talk for thirty seconds
Answer a prompt out loud, in your own words. No clinic visit, no wearable, no reading test to memorise.
02
We measure the signal
Pitch, timing and voice quality are extracted from the waveform itself. The measurement runs on our own service, not on a transcript.
03
Watch the trend
A single reading is one data point. What matters is the line it draws over weeks, and whether that line moves.
What we measure
Four measurements, taken from the waveform.
No transcript, no questionnaire, no self-report. These come out of the signal itself, which is what makes them comparable from one week to the next.
Pitch stability
jitter, %
Cycle-to-cycle variation in vocal fold vibration. Steady phonation varies little; tension and fatigue push it up.
Harmonic clarity
HNR, dB
How much of the voice is periodic tone rather than noise, measured per voiced frame from the autocorrelation peak.
Pause structure
ratio and mean length
How often you stop mid-speech and for how long, counting only gaps beyond 150 ms so stop consonants are not read as pauses.
Speech rate
words per minute
Syllable nuclei counted from the loudness envelope, then converted to words. Outside a plausible range we report nothing rather than a wrong number.
Some measures need more signal than a phone microphone can give. When a recording cannot support one — shimmer without a sustained vowel, for instance — we leave it empty instead of filling the gap with an estimate.
Why voice
Speech is the cheapest signal we have, and almost nobody is collecting it.
Producing a sentence takes precise coordination between breath, larynx and articulators. That coordination is measurable, it changes before people notice anything themselves, and capturing it costs one microphone and thirty seconds.
We are building the measurement layer first: a reading that means the same thing every time it is taken. What that reading can eventually tell you is a research question, and we would rather answer it with data we collected honestly than assume it now.
- 55M
- people living with dementia worldwide
- 10M
- new cases every year
- 7th
- leading cause of death globally
Source: World Health Organization.
Privacy
A recording of your voice is not ordinary data.
It identifies you, it carries health information, and it cannot be reissued if it leaks. We handle it accordingly, and we describe what we do in terms you can check.
Your audio has no public link
Recordings are stored under a key scoped to your account and streamed back only to you. A guessed key belonging to someone else is refused, not served.
Every request is signed
The app mints a short-lived signed token for each call to the analysis service. Nothing long-lived is ever held in your browser.
One third party, named
Measurement runs on our own service. Live conversations are streamed to Google's speech model to produce the spoken reply — that part leaves our infrastructure, and you should know it does.
Where the line is
What Harmony is not.
Voice health is a field with more confident marketing than evidence. Here is the part we would rather you read before signing up than after.
It is not a diagnosis
Nothing here is validated as a clinical instrument. A reading tells you what your voice did in those thirty seconds, and that is the whole of the claim.
It is not a screening test
Harmony does not tell you whether you have a condition, and it will not tell you that you are in the clear. If something concerns you, that conversation belongs with a clinician.
Classification is switched off
The build running today reports measurements only. Until a model is trained and validated on data we trust, it returns no classification at all rather than an unverified one.
Take the first reading today.
The line only becomes useful once it has a starting point. Thirty seconds now, thirty seconds next week.