Checking…
Checking…
Checking…
Account
1 · Enroll your voice
Read this passphrase aloud when prompted (3×):
Only the embedding (a number vector) is sent to your account on the server — never raw audio. The model itself runs 100% on-device.
2 · Log in with your voice
What I heard:
This browser has no speech recognition — after recording you can type the digits you spoke.
Type the digits you just spoke:
Attempt score: — was that really you? (Labels calibrate the match threshold.)
PIN fallback
For colds, noise, or no mic. Set while logged in, then log in any time with email + PIN.
WebAuthn
Passkey / fingerprint login is staged for a later pass — the server challenge flow above is already compatible with it, but no WebAuthn ceremony is wired up yet. Password and PIN work today.
build …
- Anti-replay: every voice login speaks fresh random digits (single-use, 2-minute expiry). A recording of a fixed passphrase will not pass.
- Honest limit: a dedicated attacker holding recordings of you speaking all ten digits could splice a challenge phrase. True audio-liveness analysis is out of scope.
- Voice login is optional — guest discovery and password/PIN login are never gated by it.
- ← Back to MUSE-AI