Barge-in

Three things must happen the moment barge-in is detected. Text-to-speech generation stops. Audio already buffered downstream — in the media server, in the carrier, in the jitter buffer — is discarded rather than played out. And the agent’s conversational state is corrected, because the caller heard only part of the sentence the agent believes it delivered in full.

The hard engineering problem is echo. The agent’s outbound audio leaks back into the inbound stream through hybrid coupling on the line or an open speakerphone, and a naive detector treats that leakage as the caller speaking, cutting the agent off mid-word for no reason. Acoustic echo cancellation, which subtracts a filtered copy of the known outbound signal from the inbound one, is what makes barge-in usable rather than a source of random self-interruption.

Barge-in and interruption handling are distinct layers and are often conflated. Barge-in is the mechanism: detect overlapping speech and stop the audio. Interruption handling is the policy: decide what the interruption meant and what the agent should say next. A system can implement barge-in flawlessly and still handle interruptions badly, by resuming a cancelled sentence from the top as though nothing happened.