BroadRouter
Get started

Voice activity detection (VAD)

Updated

Voice activity detection (VAD) is a signal-processing step that classifies short frames of audio as speech or non-speech. Speech systems use it to skip silence, split long recordings at pauses, detect when a speaker has finished, and avoid transcribing background noise.

Also called: voice activity detection, speech activity detection, speech detection, silence detection

VAD answers one question many times per second: is someone talking in this slice of audio?

How it works

Audio is cut into short frames and each frame gets a speech or non-speech label. The two most used open detectors show the typical parameters:

WebRTC VAD Silero VAD
Type classic detector from Google’s WebRTC project pre-trained neural model, MIT licence
Sample rates 8, 16, 32 or 48 kHz 8 or 16 kHz
Frame or chunk 10, 20 or 30 ms 30+ ms
Audio format 16-bit mono PCM only as documented in the repo
Tuning aggressiveness 0 (least filtering) to 3 (most) see repo examples
Speed small C code under 1 ms per chunk on one CPU thread; about 2 MB model

Frame size and sample rate are linked. At 16,000 samples per second, a 30 ms frame is 16,000 x 0.030 = 480 samples, or 960 bytes of 16-bit audio. Raw frame labels flicker, so tools smooth them: require a few hundred milliseconds of speech before opening a segment, and some silence before closing it.

Hosted APIs use the same idea. OpenAI’s speech-to-text guide lets you set chunking_strategy to auto or a VAD configuration on audio longer than 30 seconds for its diarization model, which the guide lists as deprecated.

Why it matters

  • Cost. Speech-to-text APIs usually bill per audio minute. Cutting long silences before upload cuts billed minutes.
  • Latency. Less audio uploaded means a smaller request and less to process.
  • File limits. Splitting at pauses, rather than at fixed times, avoids chopping a word in half when a file is too big for one request.
  • Voice bots. Detecting the end of a turn decides how fast a bot replies.

VAD is also the first stage of speaker diarization.

On BroadNet Inference

BroadNet’s docs do not document VAD settings or a silence-trimming option for /v1/audio/transcriptions. You upload WAV or MP3 up to 26,214,400 bytes and get back {"text": ...}. If your audio has long gaps, as voice notes and call recordings often do, run WebRTC VAD or Silero VAD locally, keep the speech regions, and send those. The WhatsApp voice notes guide covers the ffmpeg conversion to 16 kHz mono that both detectors accept. Transcription is priced per minute, $0.0042 on BroadNet Transcribe Arabic; the speech-to-text cost calculator shows what trimmed minutes save.

Questions

What is a good open-source VAD?

Two common choices are WebRTC VAD, a small classic detector with aggressiveness modes 0 to 3, and Silero VAD, an MIT-licensed neural model of about 2 MB that processes a 30+ ms chunk in under 1 ms on one CPU thread.

Does BroadNet apply VAD to my audio?

BroadNet's docs do not document VAD settings. Trim silence or split at pauses on your side before upload if you need control over it.

How does VAD cut speech-to-text cost?

Per-minute APIs bill the audio you send. If a recording is 40% silence and you send only the speech regions, you pay for about 60% of the minutes. Keep short pauses so words are not clipped.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity