Skip to content
AmplifierA man speaking on his phone

Most AI hear words.
We hear the rest.

Sona-2 · voice foundation model

Voice carries a biological signal.
We measure it.

Convert everyday voice interactions into structured health, wellness and human insights that your product can act on

2.4Munique speakers
14,000physician-labeled conditions & states
75+production indicators
40+languages

Go beyond the transcript.

A transcript tells you what a person said. Sona-2 tells you how they're doing while they said it. Same audio, same 15 seconds, one API call.

what was said

“I'm okay. We can keep going.”

transcription output · verbatim

A transcript is a record of vocabulary. It carries no information about the state of the person who produced it.

how they said itreadout · haven
  • stressEST
    elevated
  • fatigueEST
    moderate
  • mood-valenceEST
    low
  • energyEST
    below baseline
  • sleep-disruptionEMG
    moderate

Illustrative readout in Sona-2's response shape. Sample audio, not a patient recording. Screening signals, not diagnoses.

02 · two ways to run on Amplifier

One base model.
Infinitely configurable.

Use some or all of our 75+ off the shelf indicators, or build bespoke capabilities for your business problems.

base model

Sona-2™

The engine as it stands today. Established signals, one endpoint, live in weeks.

trained on
2.4M IRB-collected clinical encounters, each paired to a hand-labelled medical record
signals
75+ signals off the shelf, no configuration
hosting
Shared endpoint. First call in minutes, not a build cycle
pricing
Per assessment, from $0.10. No build fee
see the categories
your custom model

Custom inference

The same engine, fine-tuned on your voice data toward a signal you define, on an endpoint only you call.

trained on
Our corpus plus your voice data
signal
Refit an existing sign, or build a new one
hosting
Dedicated endpoint, hosted by us. Weights stay with Amplifier
pricing
One-time build fee, then per call
see pricing

Two routes in.

route a — refit

An established sign, refit to your population. Known evidence tier, low feasibility risk, fastest path to production.

discovery optional, straight to co-development
route b — new signal

A signal we have never built. Research, unproven, evidence tier starts at zero. In scope — and we say so before you spend.

mandatory paid discovery, then co-development
03 · the capability index

3 fields of voice insights. All longitudinal

01

Contextual

Age, native language, demographic characteristics. Used to sharpen the accuracy of everything else.

02

Wellness

Mood, energy, stress, sleep, hydration, cognitive load.

03

Clinical

Disease-state presence and severity across neurological, cardiometabolic, respiratory, behavioral and cognitive conditions.

04 · the Sona-2 API

Easily accessible via API & MCP

Point Sona-2 at any audio stream and get a typed signal object back. No ML team required.

  • Stream from anything: telephony, WebRTC, files, or the mic.
  • Explainable JSON, with the evidence tier on every indicator.
  • Sub-300ms round trip, streaming token by token.
  • Custom classifiers are called the same way — same endpoint, different model string.
analyze.py
response = httpx.post(    "https://api.amplifierhealth.com/v2/models/pulse/analyze",    headers={        "X-Account-ID": os.environ["AMPLIFIER_ACCOUNT_ID"],        "X-API-Key": os.environ["AMPLIFIER_API_KEY"],    },    files={"audio": ("recording.wav", audio_bytes, "audio/wav")},)
06 · built for enterprise

The audio never becomes the asset.

  • SOC 2TYPE II
  • HIPAACOMPLIANT
  • GDPRREADY

Consent first.

Sona-2 runs only on audio you are authorized to process. Consent and purpose are enforced at the API boundary, not in a policy document.

No raw audio retained.

We process in-stream and keep the readout, not the recording. Run it fully on-device when the data cannot leave the room.

Explainable by design.

Every score traces back to the moment of signal that produced it, auditable by clinicians, reviewers and regulators.

Bring us a voice stream and a question.

The first conversation is about whether your signal is real. We can usually tell you inside a week.

You bring
30 min of your existing audio
You get
a written feasibility read