Barge-in

Barge-in is the ability of a caller to interrupt a voice agent while it is speaking — and have the agent actually stop, listen, and respond to the interruption. When a caller starts talking over the agent’s audio, a barge-in-capable system detects the incoming speech, halts playback within a few hundred milliseconds, and treats the caller’s words as the new conversational turn.

Technically, barge-in requires three things working together: continuous voice activity detection running while the agent is speaking (not just while it’s listening), echo cancellation so the agent’s own outbound audio isn’t mistaken for caller speech, and a pipeline that can cancel in-flight TTS synthesis and LLM generation cleanly without leaving the conversation state confused about what was actually said versus cut off.

Why it matters for voice agents

Humans interrupt each other constantly — to correct a misunderstanding, to skip information they already know, or simply to say “yes, that one” before a list of options finishes. An agent without barge-in forces callers to sit through every sentence to completion, which is the single fastest way to make an AI feel like a 2005-era IVR (“listen carefully, as our menu options have changed”).

Barge-in quality is also a latency issue in disguise. If the agent takes 800ms to stop talking after the caller interrupts, both parties end up speaking over each other, the caller stops, the agent stops, and the conversation collapses into the awkward “no, you go ahead” loop. Good implementations stop output in under ~200ms and correctly track which portion of the agent’s response the caller actually heard, so the dialogue state stays accurate.

There’s a tuning trade-off, too: an over-sensitive barge-in triggers on background noise, coughs, or “uh-huh” backchannels, cutting the agent off mid-sentence for no reason. Production systems distinguish genuine interruptions from acknowledgment sounds before halting playback.

Related terms

• Voice Activity Detection (VAD) — the detection layer barge-in is built on
• Endpointing — deciding when the interrupting caller has finished their turn
• Time to First Audio (TTFA) — the speed metric for the agent’s reply after an interruption