Voice activity detection (VAD)
Voice activity detection (VAD) is a preprocessing step that identifies which parts of an audio stream contain human speech versus silence, music, or noise, so a transcription model only processes the spans that carry words.
Voice activity detection (VAD) scans audio and marks the spans where someone is actually speaking, separating them from silence, pauses, and background noise. In a transcription pipeline it usually runs first: the recording is split into speech segments, non-speech regions are skipped, and only meaningful audio is passed to the recogniser.
This matters for two reasons. It cuts wasted compute and shortens processing time by not decoding empty stretches, and it often improves accuracy because the model is not forced to invent words from silence or noise. VAD typically sits ahead of the automatic speech recognition (ASR) step that produces the transcript text, and it also feeds later stages such as speaker diarization and chunking. But aggressive settings can clip quiet or trailing speech, so VAD behaviour is one reason two tools differ on the same hard recording — especially with overlapping speakers, far-field microphones, or heavy background sound.