Voice activity detection (VAD)
Voice activity detection is the step that decides, frame by frame, whether an audio stream contains speech or only silence and noise.
A VAD model looks at short slices of audio, typically tens of milliseconds each, and labels each one as speech or non-speech. Everything downstream depends on that label: speech recognition only runs on the parts that contain speech, and a voice agent uses the end of speech as one signal that the caller may have finished talking.
On phone calls VAD has to cope with line noise, background voices, music on hold and the agent’s own voice echoing back. A detector that is too eager treats a cough or a passing car as speech; one that is too cautious clips the start of words or waits too long before the agent replies.
VAD is not the same as turn detection. VAD answers “is someone speaking right now?”; turn detection answers “has this person finished their turn?”, which also depends on what they said and how they said it.
