Definition
Barge-in is a caller speaking while the agent is still talking. The term also names the capability that lets the system accept that speech and stop the agent's prompt, instead of making the caller wait.
How barge-in works
The term comes from phone menu systems. VoiceXML 2.0, the W3C standard for those systems, gives each prompt a bargein attribute. When it is true, the user can interrupt the prompt. When it is false, the user must wait for the prompt to end.
The same standard defines how the interruption is detected, through bargeintype. With speech, the prompt stops as soon as speech or a keypad tone is detected, whatever was said. With hotword, the prompt keeps playing until the input matches a grammar the system is listening for, and anything else is ignored. The hotword option exists because noise, coughs and side conversations would otherwise cut prompts off.
A modern voice agent has the same decision to make, with a streaming recognizer and an LLM in place of a grammar. It must listen while it speaks, which means separating the caller’s voice from its own audio echoing back down the line. Then it must decide whether the sound is a real interruption or something to ignore, such as “mm-hm”. That last case is backchanneling.
Detecting barge-in is only the first half. Acting on it is interruption handling: the agent stops its speech, keeps track of what the caller actually heard, and responds to what they said.
Why it matters for voice agents
Callers interrupt constantly. They correct a date, say “yes, that’s right” before a read-back ends, or ask for a person halfway through a menu. An agent that cannot be barged in on forces them to listen to the whole sentence, and that sounds like a machine.
Two things go wrong in practice. An agent that is too sensitive stops for a door slam or a television in the background, then trails off mid-word. An agent that is too slow keeps talking for a second after the caller has started, so both are speaking at once.
Some content should not be interruptible. A required disclosure or a payment amount read-back is often played with barge-in off, so the caller hears all of it before the agent takes a question.
Barge-in is a caller behavior. How the system responds to it has its own name, and the two should be tested separately.
Example
An illustrative call:
- Agent: “Your appointment is on Thursday, October 8, at two thirty, with Doctor…”
- Caller: “Actually, make it Friday.” (starts speaking at “with”)
The caller’s speech is the barge-in. Whether the agent stops within a fraction of a second, drops the old date and handles “Friday” is interruption handling.