Speech recognition and models
Voice activity detection (VAD)
The technique that decides where in an audio recording somebody is speaking, and where it is silent.
Voice activity detection, abbreviated VAD, is the technique that decides where in an audio track somebody is speaking and where it is silent.
It is a preparatory step, not a result in itself. But it affects a lot of what you notice.
What it is used for
Cutting silence. A two-hour meeting recording can contain twenty minutes without speech. VAD lets the system skip them, which makes transcription faster and cheaper.
Splitting long recordings. Models process audio in chunks. VAD finds natural cut points in pauses instead of in the middle of a word.
Ending dictation. When you dictate and the system stops by itself after a few seconds of silence, it is VAD that has decided you were finished.
Controlling recording. Recording features that only capture when somebody is speaking are built on the same thing.
Why pauses disappear
A side effect worth knowing: if the silence is cut away, the information about how long the pauses were disappears with it. For a meeting summary that does not matter. For conversation analysis in research, where pause length is data, it is a loss. If you need the pauses, ask for verbatim transcription with pause marking, and check that the tool actually keeps them.
VAD also struggles with steady background noise, such as ventilation, because the boundary between speech and noise becomes less distinct.