Speaker verification

POST /v1/voiceprints/verify compares a recording with a voiceprint (stored id or token).

curl https://stt.eu.mihu.ai/v1/voiceprints/verify -H "Authorization: Bearer $KEY" \
  -F voiceprint=9b2f6c1e-3d4a-4f8b-a1c2-7e5d9f0b3a61 -F file=@call-start.wav -F nonce=session-77

{"id": "e4c1…", "object": "voiceprint.verification", "voiceprint": "9b2f6c1e-…", "score": 0.81, "match": true,
 "decision": "match", "threshold": 0.55, "speech_seconds": 6.2, "created_at": 1793171000, "nonce": "session-77", "account": "7d1e…",
 "signed": "eyJhbGciOiJFZERTQSIsInR5cCI6IkpXVCIsImtpZCI6Ij…"}

Field	Meaning
`voiceprint`	Stored voiceprint `id` or a `store=false` `token`
`file` or `file_url`	The speech to check: at least 3 s of speech (else `422`), 5 s or more recommended, at most 120 s, one voice
`nonce`	Optional (≤ 200 characters): your value (e.g. a session or challenge id), returned and signed — lets your backend refuse a replayed result

score: similarity of the two voices (about −1 to 1, higher = more alike); match = score ≥ threshold (0.55). decision gives three levels: match (≥ 0.60), uncertain (0.50–0.60: ask for more speech or another check), no_match (< 0.50).
Measured (Common Voice Turkish, 171 speakers, 20 s enrolment, 5 s test, threshold 0.55): wrongly accepted / wrongly rejected 0.3 % / 1.7 % on wideband audio, 0.8 % / 3.9 % on telephone audio, 0.7 % / 11 % on telephone audio with noise. With 3 s of speech the errors are about 2–3× higher. Real calls (other codecs, packet loss) can be worse.
A helper signal, not a login. Verification is not protected against recorded or synthetic (cloned) voices. Use it together with other checks (caller number, knowledge questions, OTP), never as the only way into an account.
Personal data: verifying a person by voice processes biometric data; your consent text must cover this purpose.
Calls are not billed.

Signed result: `signed`

signed is a JWT (JWS compact, alg: EdDSA — Ed25519) of the result, so a backend can trust a result that passed through an app or another service. Claims: iss "umay", aud (your account), sub (voiceprint id or "token"), jti (verification id), iat, exp (5 minutes later), score, match, decision, threshold, speech_seconds, nonce. The public key is a JWK set at GET /v1/voiceprints/jwks (cache it; kid names the key).

import jwt, requests                                  # PyJWT with the cryptography package
jwks = jwt.PyJWKClient("https://stt.eu.mihu.ai/v1/voiceprints/jwks", headers={"Authorization": f"Bearer {KEY}"})
key = jwks.get_signing_key_from_jwt(signed).key
claims = jwt.decode(signed, key, algorithms=["EdDSA"], audience=MY_ACCOUNT)   # checks signature, exp, aud
assert claims["nonce"] == expected_nonce and claims["match"] is True

import { createRemoteJWKSet, jwtVerify } from "jose";
const jwks = createRemoteJWKSet(new URL("https://stt.eu.mihu.ai/v1/voiceprints/jwks"), { headers: { Authorization: `Bearer ${KEY}` } });
const { payload } = await jwtVerify(signed, jwks, { algorithms: ["EdDSA"], audience: MY_ACCOUNT });

Getting Started

API Reference

Transcription

Voice

Signed result: `signed`

​Signed result: signed

Signed result: `signed`