Voice activity detection

VAD is the lowest layer of a voice pipeline. Before anything is transcribed or understood, something has to decide whether there is speech to process at all. It runs continuously and cheaply, classifying each short slice of audio as speech or not.

Everything above it depends on the answer. Endpointing uses VAD to find the gaps between utterances; turn detection uses it to notice the caller has started talking; barge-in uses it to know when to stop speaking. A VAD that mistakes a passing truck or a television for speech will make the agent stop mid-sentence for nothing; one that misses a quiet caller will make the agent talk over them.

Phone audio is narrow-band and often noisy, hands-free in a van, a busy reception, a mobile with poor signal, which is why VAD quality is a real differentiator on inbound business calls rather than a technicality.