What liveness detection is
Liveness detection establishes that the audio comes from a person speaking live, rather than a recording played back, a clone played through a speaker, or audio injected into the stream. In the security literature this family of checks is often called presentation attack detection or anti-spoofing.
It is one of the four signals in a voice identity layer, alongside speaker identity, synthetic-speech detection and the combined trust decision.
Passive vs. active liveness
Passive liveness works on whatever the person naturally says. Nothing changes for the user: the model analyses the audio in the background and returns a score. This is the right default for voice agents and robots, where an extra step would break the conversation.
Active liveness adds a challenge — for example asking the user to repeat a random phrase. It is useful as a step-up when risk is high, because a pre-recorded replay cannot answer an unpredictable prompt. The cost is friction, so it is best reserved for the moments that need it.
What the model listens for
Live speech captured directly by a microphone differs from speech that has been recorded, stored and played back. Playback adds the response of a loudspeaker, the acoustics of a second room and the characteristics of a second capture. Synthetic speech adds artefacts of its own generation. A liveness model learns these differences from large amounts of genuine and attack audio.
Context also matters. Vosco classifies the audio channel — studio, phone or noisy — so a verdict on compressed telephony audio can be interpreted differently from one on a clean device microphone.
Liveness and speaker verification work together
A voiceprint match on its own cannot tell a live customer from a recording of that customer. Liveness on its own cannot tell the right live person from the wrong one. Combined, they answer the question that matters: is the enrolled person speaking right now? The difference is covered in more depth in replay attacks vs. synthetic voices.
Liveness on the device
For robots and smart devices, liveness has to run locally — a recorded command played from a phone carries the same authority as the owner who recorded it. Vosco Nano runs liveness, replay, synthetic and speaker signals on the chip, fully offline, in under 10 ms on an NPU. See Vosco for Physical AI.
Frequently asked questions
What is voice liveness detection?
Voice liveness detection checks that a real person is speaking live into the microphone, rather than a recording, replay or synthetic voice being played back.
What is the difference between passive and active liveness?
Passive liveness analyses natural speech with no extra steps for the user. Active liveness asks the user to respond to a prompt, which adds friction but defeats pre-recorded audio.
Do I still need speaker verification if I have liveness?
Yes. Liveness confirms the speech is live; speaker verification confirms who is speaking. Together they confirm the enrolled person is speaking right now.
Verify the voice before your AI acts.
Start on the Vosco API with 1,000 minutes free every month, or talk to us about on-device deployment.