hotai.news
← Back to AI WikiAI Wiki / Core concepts

What is voice activity detection? Separating speech regions from silence

Understand VAD, ASR, and diarization through a long meeting, including thresholds, pauses, clipping, quiet speech, and global timing.

HotAI.news · Sources checked with CodexPublished on 2026-10-20Sources checked on 2026-10-20
Our take

Understand the concept

VAD estimates speech regions for segmentation or live-turn management. It does not transcribe content or identify speakers. Noise, quiet speech, thresholds, and pauses can cause misses or clipping. Preserve timing and inspect boundaries rather than assume nonspeech classifications contain nothing important.

01

Speech intervals differ from transcription output

A meeting’s quiet stretches can be segmented before ASR, but VAD does not create its transcript. Silero provides detectors and Google exposes speech events with implementation-specific definitions.

Music and laughter can trigger errors, and distant quiet speech can be missed. A colored interval plot does not prove complete coverage. No meeting audio or universal noise-robustness test was performed.

02

Thresholds and pauses balance missed speech and fragmentation

Stricter detection can remove noise and quiet word endings. Short-pause endpointing fragments speech, while long waits affect interaction. Check word boundaries and continuous speech rather than silence-removal percentage alone.

Padding can help boundaries, with task-specific values. Google’s activity events and timeouts also affect stream control; end-of-speech is not receipt of all finalized transcription results.

03

Preserve original timing after segmentation

After removing silence, captions need a clearly chosen original or edited timeline. Keep segment start offsets for global times. This is design without audio trimming; inspect boundaries and disputed regions.

VAD can reduce later work without replacing ASR or diarization evaluation. Nonspeech sounds may matter. Document removed segments and uncertain boundaries rather than treat preprocessing as completed transcription.

Checked sources

These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.

Back to AI Wiki ↗