Voice activity detection. VAD, is the layer that decides, in real time, whether the audio coming in is the caller speaking or silence. It is the basis for endpointing (when the caller is done), barge-in (when the caller cuts in on the agent), and turn-taking (whose turn it is to speak).
Cheap VAD trips on background noise, breath sounds, and mid-sentence pauses. The result is an agent that interrupts the caller mid-thought, or sits silent when they finish. Both feel broken.
Production-grade VAD is sensitive to the caller's speaking pace (which is vertical-specific, clinic patients pause more than auto-shop callers asking about brakes), the noise profile of the environment, and acoustic markers of incomplete thoughts. A serious vendor tunes VAD per-vertical, not globally.
Vorel ships Bayesian VAD tuned per-vertical, the clinic profile is more patient than the dispatch profile, by design.