Why matching isn't enough
Biometric systems answer "is this the right face/voice?" Liveness answers the prior question: "is this a real face/voice, right now?" Without it, a printed photo, a screen replay, or a cloned voice can pass a matching check perfectly — because it genuinely does match.
Active vs passive liveness
| Active liveness | Passive liveness | |
|---|---|---|
| Method | Challenge–response: blink, turn your head, read a phrase | Analyzes the sample itself: texture, depth, micro-movement, spectral properties |
| User friction | Higher — requires cooperation | None — invisible to the user |
| Replay resistance | Strong (response must match a fresh challenge) | Good and improving |
| Typical use | High-risk onboarding, step-up checks | Continuous or high-volume verification |
Voice liveness: the neglected channel
Almost all liveness investment has gone into faces, driven by KYC onboarding. But the highest-value fraud in 2026 happens on calls: a cloned executive voice requesting a wire. Voice liveness applies the same two families of technique — passive acoustic analysis of the live audio, and active challenges (unexpected questions, mid-call verification phrases) that pre-generated or streaming clones handle poorly.
The practical ceiling is the same as any detection layer: it estimates, it doesn't guarantee. That's why voice liveness belongs in front of a deterministic control — callback verification — rather than replacing it. Score the call; verify the payment.