A phone conversation is not a polite relay. People correct their name mid-sentence, say “yes, yes” before the question ends, remember something else and talk over you. A real person on the line stops, listens and picks up again. A voice agent that keeps talking to the end of its sentence makes it obvious, within two seconds, that a machine is reading a script on the other end.
Barge-in is a voice agent’s ability to stop talking when the person starts speaking over it. In practice it takes three things: noticing the user’s speech while the agent’s audio is still playing, discarding the audio already queued to go out, and recording in the history only the part the caller actually heard. Without the third, the agent “remembers” saying things nobody heard.
Why stopping is harder than it looks
In a chained voice pipeline (transcription, language model, speech synthesis), the reply becomes audio before it reaches the caller’s ear, and part of that audio is already sitting in the line’s buffer. Stopping generation is not enough: the telephony provider has to be told to drop whatever has not played yet. And if the system only notices the interruption when the final transcript arrives, the caller has already spoken a whole sentence over a voice that would not stop.
False positives are the enemy
The caller’s microphone picks up everything: a cough, an “uh-huh”, the TV in the background and the echo of the agent’s own voice coming back down the line. An agent that stops at every noise stutters and never finishes a sentence. That is why the “started speaking” signal cannot be the trigger on its own.
- Voice activity detection marks a candidate interruption, but does not interrupt yet.
- The interruption only fires when the partial transcript contains real words, two or more, after a minimum amount of speech.
- Right after the agent starts talking there is a short echo-guard window, because that is exactly when its own voice comes back down the line.
- A cap on interruptions per minute keeps a noisy line from turning the conversation into a loop of half-started sentences.
An agent that never stops sounds like a recording. An agent that stops at any noise sounds worse: it sounds like the call is breaking up.
The history has to reflect what was heard
This is the detail most often missed. If the agent was about to explain three plans and got interrupted on the first, the conversation record cannot contain all three. Otherwise, on the next turn, the model behaves as if the caller already knows the other two, and the summary that reaches the sales team describes a conversation that never happened. The right move is to cut the text at roughly the point where the audio stopped and carry on from there.
What an interruption does not undo
Stopping speech is safe. Undoing an action is not. If the agent was querying a system or placing an order when the caller cut in, the audio stops but the action runs to completion. Design the agent’s tools with that in mind: confirm with the caller before doing anything irreversible, not after.
How it works at Take Sales
Take Sales voice agents run in two modes. In native real-time audio mode, the model itself handles interruptions. In pipeline mode, which combines separate transcription, language model and speech synthesis, barge-in has been live in production since September 2026 with the safeguards above: a transcript-confirmed trigger, echo guard, a per-minute cap and history cut at the spoken part. It applies to voice in the website widget and to phone calls alike.
The thirty-second test still holds: talk over the agent while it is answering. The piece on latency explains why that test says so much about the architecture, and the IVR comparison shows what changes for the caller.
