Gatekeeper of the pipeline. Bad VAD shows up as the agent talking over people or waiting too long before replying.
- Works on short audio frames and decides whether each frame contains speech. Adjacent speech frames are grouped into a segment before going downstream.
- A simple detector can use short-time energy and zero-crossing rate. Neural detectors handle messy rooms better, but cost more and can still fail on music, crosstalk and breathing.
- Quality is not just classification accuracy. In an agent, the visible failures are clipped first words, long waits, and the agent speaking over the caller.
Connections
- precedes Speech-to-Text
- part_of Voice Agent