NewCall a live VoiFlow agent in your region. It picks up on the first ring.Call it now
VoiFlow
All resources

Technology 5 min read

Barge-in and turn-taking: teaching an AI agent when to stop talking

Barge-in is when a caller talks over the agent. How to stop quickly, ignore backchannels, tell when the caller has finished and test it on real lines.

Engineering
VoiFlow

Two colleagues talk at a café table, one mid-sentence with a hand raised while the other leans in to reply

Callers interrupt. They say "no, wait" halfway through a sentence, or cut in to correct a date they have just heard. Barge-in voice AI is the ability to stop speaking when a caller interrupts, listen, and then answer what was actually said. It is one of the clearest signs of whether an agent is a good listener.

Barge-in is a caller speaking over an agent while it is talking. A good agent stops quickly, listens and answers the point the caller made. It should not stop for a short "mm-hm" or a cough. Telling a real interruption from a sign of listening, and deciding when the caller has finished, are the two hard parts.

What happens when a caller interrupts?

Barge-in starts while the agent is still speaking and the line is still listening. Voice activity detection, a component that separates speech from silence and noise, picks up the caller's voice. The agent then stops its own audio. That has to happen fast enough that the caller does not hear a long overlap. The caller's words then go to speech recognition, and the agent decides what to do with them.

The agent also has to avoid hearing itself. When the caller plays the agent's voice through a speaker, the agent's own speech can loop back into the line. Echo cancellation, a signal step that removes the agent's own voice from what it hears, prevents the agent from thinking the caller is interrupting when it is only hearing its own words. This is part of the wider flow described in how AI voice agents work.

Which interruptions should stop the agent?

Not every sound from the caller is an interruption. Short signs that someone is listening, such as "mm-hm", are called backchannels. The agent should sort what it hears into a few groups, and respond differently to each.

What the caller saysWhat it usually meansWhat the agent should do
"Mm-hm", "yeah", "right"A sign of listeningKeep talking
"Sorry, which day?"A clarifying questionStop and answer it
"No, it's the fourteenth"A correctionStop, take the correction and confirm it
Someone else speaking in the roomNot addressed to the agentKeep talking, unless the caller answers them
A cough, a door, a car hornNoiseIgnore it, or pause briefly and resume

Start with the groups that matter most for your calls. A booking line gets many corrections to dates and times, while a payment reminder line gets many questions about amounts.

These groups are judgements about meaning. A timing rule alone cannot tell a correction from a "yeah". Decisions about meaning should come from the language model reading the words, not from a list of keywords, with timing rules doing only the first filtering.

How does the agent know the caller has finished?

This is called endpointing: deciding that a turn has ended. The simplest method waits for silence of a set length. Set that too short and the agent cuts the caller off while they think. Set it too long and the agent sits in silence after every question, which is the same delay that makes calls feel slow. The cost of that delay is covered in voice AI latency.

Good systems also look at the words. A caller who says "I'd like to book for Tuesday, and" has not finished, even after a short pause. A timing rule and the model's reading of the sentence together work better than either alone.

Background noise and speakerphones

Most barge-in failures happen on imperfect lines rather than in a quiet office. Four problems come up again and again.

  • Speakerphones. The agent's voice comes out of the caller's speaker and back into the microphone. Echo cancellation and sensible output volume both help.
  • Cars, cafés and open offices. Background noise can look like speech. Noise suppression removes steady sound, but it can also clip quiet speech, so test both.
  • Compressed phone lines. Compression can blur the sounds that separate similar words. Callers may need to repeat themselves, which is acceptable. Guessing is not.
  • Soft-spoken callers. A low threshold picks up background noise, and a high one misses soft speech. Tune it on real calls rather than in the office.

When the audio is unclear, the agent should say so and ask the caller to repeat the point. It should not guess and carry on.

How do you test barge-in before go-live?

A small, structured test will show most problems. Build it from real calls rather than invented ones.

  1. Write down 30 interruptions taken from recent calls: backchannels, corrections, cut-ins and noise.
  2. Play or speak each one on several phone types: mobile, landline and speakerphone.
  3. Time the gap from the start of the interruption to the agent stopping. Note any case where it keeps talking too long.
  4. Check what the agent says next. Did it answer the point, or repeat itself?
  5. Run the same set again after any change to the voice, the model or the timing settings.

Simulated checks can confirm the logic, but only real phone calls show how a line behaves. Run the set on live calls before you rely on the results.

What should you check in a demo?

  • Does the agent stop within a moment of being interrupted, on a phone line and not only in the demo room?
  • Does it keep going through "mm-hm" and similar sounds?
  • After being interrupted, does it answer the point the caller made?
  • Can you replay the call and see exactly where it stopped and why?

The first seconds of a call set the tone for everything after them, and the first ten seconds of a call covers how to get them right.

How VoiFlow handles this

Every VoiFlow call is recorded, transcribed, scored and replayable, so the moments an agent talked over a caller can be found and fixed. Staff can listen in on a live call, whisper to the agent and take over if an interruption goes wrong. The conversation platform page describes how turns are managed. Ask any vendor to show a replay of a barge-in case, including what the agent heard and when it stopped.

Frequently asked questions

What is barge-in in voice AI?

Barge-in is when a caller speaks while the agent is still talking. A well-built agent stops, listens and answers what the caller said, rather than finishing its own sentence.

Should an agent stop for every sound the caller makes?

No. Sounds such as "mm-hm" or a cough are not interruptions. The agent should keep talking through backchannels and ignore noise, and stop for questions, corrections and requests.

How long should an agent wait before it replies?

There is no single correct pause. Tune the wait on real calls by call type. A short pause suits a yes or no answer, while a longer one suits a caller who is reading out a number. Check the cut-offs and the silences every week.

Can I test barge-in in a demo?

A demo shows one line in a quiet room, so it proves little on its own. Ask for a test on your own phone lines, using interruptions from your own callers. You can try a live call now, or talk to us about a test on your phone setup.

Call it now · United States line+1 (831) 282-9226

Hear it run one of your calls.

Call the agent yourself, start a free trial, or bring us your hardest workflow. Most teams are live in days.