01 — Teach it a voice
Enroll
Record up to 15 seconds under a name. Recording the same name again adds another sample and sharpens the match. Speak for at least 3 seconds.
02 — Ask who is speaking
Identify
Record a fresh sample. Every enrolled voice gets a probability; the raw cosine similarity decides whether anyone matches at all.
Probability the recording is AI-generated. Indicative demo signal, not forensic proof.
03 — Watch it decide
Live mode
Once a second the last 3 seconds and the last 10 seconds are each scored on their own. The short window reacts to a speaker change within a tick or two; the long one is steadier. During silence the lines wobble and the badge reads unknown — that is the model telling the truth, not a fault.
| Speaker | 3s | 10s |
|---|---|---|
| Start listening to fill this table. | ||