Glossary
Voice Activity Detection (VAD) is the pre-processing step that finds which parts of an audio recording contain speech and which don't. It's a binary classifier per audio frame — "speech" or "not speech". VAD runs before ASR to skip silence (saving compute) and before diarization to find the segments worth clustering.
What VAD detects
Speech vs everything else. "Everything else" includes silence, music, traffic noise, applause, laughter, mouse clicks, breath sounds, and crosstalk. A good VAD model has high recall on quiet speech (whispering, faraway voices) without false-positives on noise that has speech-like characteristics (TV audio, distant conversations).
Why VAD matters
Three reasons. (1) Compute: ASR is expensive; not running it on 30 minutes of silence in a 1-hour recording saves real money. (2) Quality: Whisper hallucinates on long silences — it'll invent plausible-sounding text. VAD prevents this by simply not sending silence to the model. (3) Diarization: clustering speaker embeddings only makes sense on actual speech — VAD gates the segments.
Common VAD models
Silero VAD — small (1.8MB), fast, runs on CPU, great quality. Open source. The de-facto choice for most production pipelines. WebRTC VAD — older, very fast (microseconds per frame), but less accurate on noisy audio. pyannote VAD — heavier, slower, but state-of-the-art on hard cases. We use Silero in our pipeline as a cheap pre-filter before sending audio to Whisper.
In practice
Drop a 60-minute Zoom recording where 25 minutes are silence (someone muted, screen-share monologue, etc.). Without VAD, we'd send the full 60 minutes to Whisper. With VAD, we only send the 35 minutes of actual speech — 40% compute savings and the transcript doesn't contain hallucinated text from silence segments.
Related terms
Further reading
Try meeting transcription
Get started — freeLast updated: September 20, 2026
Frequently Asked Questions
What does VAD (Voice Activity Detection) mean in transcription?
A pre-processing step that detects which parts of an audio stream contain speech vs silence vs noise. Saves processing time by skipping non-speech segments.
Why does VAD (Voice Activity Detection) matter when choosing a transcription workflow?
VAD (Voice Activity Detection) affects how you understand transcript quality, timing, compatibility, or the technology behind speech-to-text results.
Where can I apply VAD (Voice Activity Detection) on Transcript.you?
Use the related workflow at /zoom-to-text when you want to see how this glossary concept connects to an actual transcription task.
Browse the full transcription glossary or read the complete guide to AI transcription.