How full-duplex voice works
Half-duplex systems, like a walkie-talkie or a classic phone menu, take turns: one side talks while the other waits. Full-duplex voice AI streams the caller’s audio continuously, uses voice activity detection to notice when the caller starts speaking, and cuts its own speech short when that happens. This is often called barge-in.
The hard part is telling real interruptions from noise. A cough, a dog barking, or line static shouldn’t stop the assistant mid-word, while “wait, it’s Thursday, not Tuesday” should. Systems tune how long a pause counts as the end of a turn and how loud speech must be before it interrupts.
Example: a restaurant reservation change
A caller books a table for four at a restaurant. While the assistant reads back “four people on Friday at seven,” the caller interrupts: “Actually make that six.” The assistant stops, confirms six, checks the time still works, and finishes the booking without making the caller wait for the full read-back. This scenario is illustrative.
Interruptions in InsertChat calls
InsertChat’s AI receptionist is built for natural turn-taking. When a caller asks a question or corrects it, even during the greeting, it stops speaking, listens, and answers the latest question first. Coughs, background voices, and line noise are not treated as interruptions, and an interruption alone doesn’t cancel a task in progress.
It follows a caller’s spoken corrections for the rest of the call, including how to pronounce a name. See how it handles bookings for restaurants.
