Voice activity detection (VAD)

Voice activity detection classifies short slices of audio as speech or not speech. It's a small, fast component that sits in front of heavier models. In a voice assistant it decides when the user has finished talking; in transcription it can drop long silences and split recordings at natural pauses.

In realtime voice features, VAD settings shape how the conversation feels. Basic VAD ends the user's turn after a fixed stretch of silence, which cuts off people who pause to think, while a long silence window makes the assistant feel sluggish. Semantic turn detection, offered by some realtime APIs, uses a model to judge whether the user sounds finished instead of counting milliseconds. VAD also drives barge-in: when the user starts talking over the assistant, playback has to stop immediately.

Background noise, other people's voices, and echo from the device's own speaker can all register as speech and trigger false turns. Echo cancellation, a push-to-talk option for noisy settings, and testing with slow and non-native speakers catch most of these problems. Make the state visible, listening versus thinking versus speaking, so users understand what the system heard.