# What is Voice Activity Detection?

Also called VAD, speech detection.

Voice activity detection is the frame-by-frame classification of an audio stream into speech and non-speech. It runs ahead of transcription and other processing so that silence, noise, and background sound can be skipped. VAD supplies the raw speech and silence signal that endpointing, barge-in handling, and bandwidth control all build on.

Classic VAD used signal statistics: short-term energy, zero-crossing rate, and spectral flatness compared against an adaptive noise floor. These methods are cheap and still appear in embedded audio paths, but they confuse speech with any loud sound and degrade sharply in cars, cafes, and open offices. Neural detectors trained on labeled speech corpora now dominate, classifying short frames with far better noise robustness at modest compute cost.

The output is a stream of speech probabilities per frame, usually smoothed with hangover logic so a brief unvoiced consonant does not register as the end of speech. Systems expose tunable parameters for activation threshold, minimum speech duration, and minimum silence duration, and these values effectively set the personality of the surrounding voice interface.

VAD serves several purposes at once. It gates the transcription stream so a recognizer is not billed or burdened with silence, it drives comfort noise and discontinuous transmission in telephony codecs to save bandwidth, and it tells a speaking agent that the user has started talking so barge-in can trigger. Each consumer wants slightly different sensitivity, so a single detector often feeds several thresholds.

VAD is frequently confused with endpointing. VAD answers whether sound is speech right now; endpointing answers whether the speaker is done. Endpointers consume VAD output but add timing rules and, increasingly, language-level judgment on top of it.

## Key points

- Classifies each short audio frame as speech or non-speech
- Neural detectors outperform energy thresholds in noise
- Hangover smoothing prevents false stops on brief pauses
- Feeds endpointing, barge-in, and bandwidth savings
- Distinct from endpointing, which judges turn completion

## In practice

A browser-based voice agent runs a small detector on the microphone stream before opening a network connection. Keyboard clatter and a hallway conversation stay below the speech threshold, so nothing is transmitted. When the user speaks, the detector crosses its threshold within a few frames, the client opens a streaming transcription session, and the same signal tells the agent to stop its current spoken reply.

## Related terms

- [Endpointing](/en/glossary/endpointing)
- [Barge In](/en/glossary/barge-in)
- [Speech to Text](/en/glossary/speech-to-text)
- [Voice Agent](/en/glossary/voice-agent)
- [Real-Time Transcription](/en/glossary/real-time-transcription)

[Back to the AI Glossary](/en/glossary)