Skip to main content
VoiceMark turns any microphone into a biometric gate. Enrol a voice once, then issue short-lived JWT tokens on match — usable anywhere you currently check a password or OTP.

​
How it works

1

Enrol

Record 2–10 seconds of speech. VoiceMark extracts a 192-dim ECAPA-TDNN speaker embedding and returns a voiceprint_id. No audio is stored after processing.
2

Challenge

Request a phrase from the verse corpus. Show it to the user before they record. Prevents replay attacks.
3

Verify

Record the phrase. VoiceMark runs speaker match, returns an ephemeral JWT (60s TTL) and both VoiceMark visual outputs.
4

Authorise

Your backend validates the token at /v1/token/{id}/validate before any privileged action.

​
Key properties

  • Zero PII stored — raw audio is deleted immediately after embedding extraction
  • Ephemeral tokens — 60-second TTL, single-use
  • EER 0.0% — separation gap 0.701 at threshold 0.842 (see EER results)
  • Two visual outputs — Cloud VoiceMark (live acoustic) + Glyph VoiceMark (stable identity)

​
Base URL

All requests require Authorization: Bearer <api-key>.