Defeats replay & cloned-voice attacks that pure detection can't
Prove you're a live human
The system speaks a random number and you must say it back — live. A recording or an AI clone can't produce digits generated one second ago. Verification checks three things at once.
Content
did you say the right digits? (ASR)
Liveness
is the voice a live human? (5-model)
Timing
answered inside the window?
Allow the microphone when prompted. You'll get ~5 seconds to read the number aloud.
Get ready…
— — —
Recording — say it now · 5s
◠ Verifying content · liveness · timing…
—
We heard: · expected · synthetic score
Why this beats pure detection
Playing a cloned voice through a speaker can launder the digital artifacts that detection relies on. But an attacker with a pre-recorded or cloned sample cannot answer a challenge that didn't exist when they recorded it. Combined with real-time synthetic-speech scoring and a strict time window, this closes the replay gap — the bank's step-up path when a call scores MEDIUM/HIGH.