Reference voiceprints

Off by default and only for the requests that ask for it. Store the voice of the person you want to follow once (e.g. the account holder, with their consent), then name it on a request with reference_voiceprint=<id>. Without a reference, speaker_lock, voice_isolation and diarize take the first voice they hear as S1; with one, S1 is the stored person from the start, so a call where someone else speaks first, or a background voice that is louder at the start, is still locked to the right person.

Create: `POST /v1/voiceprints`

curl https://stt.eu.mihu.ai/v1/voiceprints -H "Authorization: Bearer $KEY" \
  -F file=@customer-1042.wav -F name="Customer 1042" -F consent=true -F consent_reference=CRM-CONSENT-88123

Field	Meaning
`file` or `file_url`	3–120 s of one person speaking (any common format). At least 3 s of speech after silence is removed; 15–20 s of speech recommended (10 s works, but verification errors are 30–50 % higher)
`name`	Your label, 1–100 characters
`consent`	Must be `true`: you confirm the person agreed to their voiceprint being stored and used. Without it → `400`
`consent_reference`	Optional (≤ 200 characters): your reference to the consent record (stored and returned as is)
`store`	`true` (default): we keep the voiceprint and answer its `id`. `false`: we keep nothing and answer a `token`

{"id": "9b2f6c1e-3d4a-4f8b-a1c2-7e5d9f0b3a61", "object": "voiceprint", "name": "Customer 1042", "speech_seconds": 14.2,
 "created_at": 1793170800, "consent": {"given": true, "at": 1793170800, "reference": "CRM-CONSENT-88123"}}

The voiceprint covers both channels: it is computed from the recording as sent and from the same speech through a simulated telephone line, so a voice recorded on a microphone or app is also recognised on a phone call.
Only the speaker embedding (a list of numbers from which the audio cannot be rebuilt) is stored, with the name and the consent record. The recording is not kept: it is deleted as soon as the voiceprint is computed. Voiceprints are stored on the server that answers the API.
Errors: 400 missing consent / name, longer than 120 s; 422 less than 3 s of speech, or more than one voice in the recording; 409 limit of 1 000 voiceprints per account reached.
A voiceprint belongs to your account (API key’s tenant); other accounts get 404.

List, read, delete

curl https://stt.eu.mihu.ai/v1/voiceprints -H "Authorization: Bearer $KEY"          # {"object": "list", "data": [...]}
curl https://stt.eu.mihu.ai/v1/voiceprints/$ID -H "Authorization: Bearer $KEY"
curl -X DELETE https://stt.eu.mihu.ai/v1/voiceprints/$ID -H "Authorization: Bearer $KEY"   # {"id": ..., "deleted": true}

DELETE removes the voiceprint for good (use it when the person withdraws consent); requests that still name it get 404. The embedding itself is never returned. Voiceprint calls are not billed.

Use: `reference_voiceprint`

# file: only the stored person's speech becomes text, everyone else comes as "ignored"
curl https://stt.eu.mihu.ai/v1/audio/transcriptions -H "Authorization: Bearer $KEY" \
  -F file=@call.wav -F model=umay -F language=tr -F speaker_lock=true -F voice_isolation=true \
  -F reference_voiceprint=9b2f6c1e-3d4a-4f8b-a1c2-7e5d9f0b3a61

wss://stt.eu.mihu.ai/v1/stream?model=umay-stream&language=tr&api_key=<KEY>&voice_isolation=true&speaker_lock=true&reference_voiceprint=9b2f6c1e-…

Works on files (sync and asynchronous; also in extra_body with OpenAI model names), /v1/stream (query) and /v1/realtime (query or session field reference_voiceprint).
Needs voice_isolation, speaker_lock or diarize (else 400). Unknown or deleted id → 404 (live: close 4400).
voice_isolation + reference: the isolation model follows the stored person from the first second; it still switches on only when another voice is heard. diarize + reference: S1 = the stored person, others S2, S3, ….
After an update of the voice model a voiceprint may answer 409 (“create it again”); this is announced in the changelog.

Not stored: `store=false`

With store=false nothing is kept on our side. The answer carries a token instead of an id: the voiceprint, encrypted and authenticated with a server key (AES-256-GCM) and bound to your account. You keep it (e.g. in your CRM next to the customer) and send it wherever an id is accepted: reference_voiceprint=<token> and POST /v1/voiceprints/verify with voiceprint=<token>.

{"id": null, "object": "voiceprint", "stored": false, "token": "vpt1.3fa1c2d9.Qm9vb…", "name": "Customer 1042",
 "speech_seconds": 14.2, "created_at": 1793170800, "account": "7d1e…", "consent": {"given": true, …}}

The token cannot be read or changed by anyone (a changed token, or one of another account → 404). It is still a derivative of a person’s voice: store it like other personal data, delete it when consent is withdrawn (that is the deletion — there is nothing on our side).
It does not show in GET /v1/voiceprints and does not count against the 1 000 limit. It stays valid until the voice model changes (then 409, create it again; announced in the changelog).

Getting Started

API Reference

Transcription

Voice

Create: `POST /v1/voiceprints`

List, read, delete

Use: `reference_voiceprint`

Not stored: `store=false`

​Create: POST /v1/voiceprints

​List, read, delete

​Use: reference_voiceprint

​Not stored: store=false

Create: `POST /v1/voiceprints`

List, read, delete

Use: `reference_voiceprint`

Not stored: `store=false`