Voice activity detection (VAD)

What it is

Voice activity detection runs on every few milliseconds of audio and answers one question: is someone speaking? It is cheap, fast and sits in front of the heavier components. Speech recognition waits for it, endpointing counts silence with it, and barge-in relies on it to tell a caller's interruption from room noise.

Simple detectors use energy thresholds and frequency patterns. Modern ones are small neural models that are more robust to noise. None is perfect: hold music, a television in the background, a carrier's comfort noise or a caller breathing close to the handset can all register as speech, and a soft-spoken caller on a poor line can fail to.

Why it matters when buying

Most turn-taking failures trace back to this component. If the detector never sees silence on a noisy line, endpointing never fires and the agent seems to ignore the caller. If it fires on noise, the agent interrupts or "hears" an empty turn and may fabricate a transcript for it. Voice agents also send audio upstream only when the detector says so in some architectures, which means a mis-tuned detector clips the first syllable of every answer. That is one reason short answers such as "yes" and "no" go missing.

What to ask

Ask how the detector is tuned for telephone audio and whether sensitivity adapts per call. Ask what the agent does when it receives a "turn" containing no words. In the demo, play music or speak from across the room while the agent talks, and give short one-word answers quickly after a question.