Barge-in

People interrupt. They answer a question before it is finished, correct a detail, or say "actually, I need to reschedule" halfway through a confirmation. A voice agent without barge-in keeps talking, the caller repeats themselves, and the conversation stalls. With barge-in, the agent detects speech, cuts its own audio, and processes what the caller said.

How it works. The telephony layer runs voice activity detection on the caller's audio while the agent is speaking. When speech is detected above a threshold, playback stops and the streaming recogniser takes over. Echo cancellation matters here: the agent's own voice coming back through the line must not be mistaken for the caller.

Tuning it. Too sensitive and background noise or a cough cuts the agent off; too lax and callers talk over a reply that will not stop. Good systems distinguish short acknowledgements ("yes", "okay") from real interruptions and resume or restart accordingly.

Why it matters for outcomes. Interruptions are where corrections happen: a wrong address, a different time, a change of mind. Agents that handle them capture accurate data; agents that do not produce bookings on the wrong day.

Example. The agent begins "So that's Thursday at 2 pm at 42 Elm" and the caller says "no, 24 Elm". The agent stops, confirms 24 Elm Street and continues. Botcadence voice agents support barge-in on inbound and outbound calls; the engineering notes on LiveKit echo cancellation cover the audio side. Botcadence voice agents meter this way on every plan.