Skip to content

Bananaflow MCP is live.

Read the docs

Barge-In: Letting Callers Interrupt AI Agents

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Cover for "Barge-In: Letting Callers Interrupt AI Agents": a white line diagram of a caller cutting in: a panel with a speaker labelled Agent sends a long bold wave to the right, a short wave from a person at the upper right bends down into a lit node, the agent's wave is cut by a short upright bar just before the node, and one arrow leads on from the node to a panel with a headset labelled Listens, over a pill reading Caller cuts in, drawn over a soft, out-of-focus cyclamen in muted rose, its slender swept-back petals streaming sideways, with a cool sky blue haze at the left end and a warm peach glow at the right.

Barge-in is when a caller starts speaking while a voice agent is still talking, and the agent stops to listen. This guide explains how barge-in works on an AI voice agent, how the agent tells a real interruption from a polite "haan", what should happen to the half-spoken sentence, and how to set and test it for Indian callers. A note on the word: in a call centre, "barge" can also mean a supervisor joining a live call. That is a different thing, covered in our glossary entry on call barging.

Here is a call that goes wrong in both directions. A bank's agent is reading a loan reminder to Mr Iyer in Chennai: "Your EMI of Rs 12,400 is due on the fifth, and a late fee of..." Mr Iyer says "haan, haan" to show he is listening. The agent stops dead and asks, "Sorry, did you want to say something?" Later in the same call he says, "Ruko, woh date galat hai," and the agent talks on for another six seconds before it hears him. Same agent, same caller, two failures. The bank, Mr Iyer and the amounts are made up for this example.

What is barge-in on a voice agent?

Barge-in is the caller speaking over the agent and the agent yielding the floor. When it works, the agent stops, listens, and answers what the caller just said. When it is switched off, the caller has to wait for the agent to finish.

The idea is old. It comes from phone menus, where a caller who already knew the option wanted to press it early. The W3C's VoiceXML 2.0 specification, the standard for such menus, describes it this way: if the platform supports bargein, the author can specify whether a user can interrupt a prompt using speech or keypad input, and it "speeds up conversations, but is not always desired" (W3C VoiceXML 2.0, section 4.1.5, read 16 November 2026).

An AI agent talks much more than a phone menu does, so barge-in matters more. A menu speaks for four seconds. An agent explaining a policy can speak for twenty.

This post is about the caller cutting in while the agent talks. The related question, how the agent knows the caller has finished, is a separate job. We cover it in our guide to endpointing and turn detection. For the short definition, see our glossary entry on barge-in.

One agent sentence on a timeline. The caller says “haan” midway and the agent keeps going. Later the caller says “ruko”, the audio is cut off, the rest of the sentence is dropped, and the agent listens.

How does a voice agent detect a barge-in?

A voice agent detects barge-in by listening to the caller's line the whole time it speaks, and combining up to three signals: sound, length and words. Using only the first is fast but jumpy. Using all three is slower but steadier.

Sound. Voice activity detection marks short slices of audio as speech or not speech. It answers "is someone making human sound right now?" and nothing more. A horn, a cough and a real word can all look like speech to it unless it is set well. Our hands-on guide to Silero VAD on 8 kHz phone audio shows the settings that matter.

Length. A real interruption usually lasts longer than a click or a cough. A minimum speech length, often a few hundred milliseconds, ignores sounds shorter than that. The trade is simple. A longer minimum ignores more noise, but the agent keeps talking over a caller for a moment longer.

Words. Once the speech recogniser has the first word or two, the agent can ask whether they are a stop word or an acknowledgement. This is the most accurate signal and the slowest, because words take time to recognise.

The W3C VoiceXML specification names the same trade. It defines two kinds of barge-in. With the "speech" type, the prompt is stopped as soon as speech is detected, whatever was said. With the "hotword" type, the prompt is not stopped until the input completely matches a grammar the platform is listening for, and other input is ignored (W3C VoiceXML 2.0, section 4.1.5.1, read 16 November 2026). Today's voice agents use the same two ideas: stop on any sound, or stop only on words that count.

  • Signal: Sound (voice activity detection)

    What it answers
    Is someone making speech-like sound?
    Speed
    Fastest
    What goes wrong
    Stops for a horn, a TV or a cough
  • Signal: Minimum speech length

    What it answers
    Has the sound lasted long enough to be a word?
    Speed
    A few hundred milliseconds
    What goes wrong
    Too long: the agent talks over the caller
  • Signal: Words

    What it answers
    Is it a stop word, or just "haan"?
    Speed
    Slowest, needs the first word recognised
    What goes wrong
    A missing stop word, so a real "ruko" is ignored

What happens to the agent's audio when a caller cuts in?

The agent's audio stops, the part it had not yet said is thrown away, and the agent answers the caller's new sentence. Three things have to happen together, and each one fails in its own way.

  1. Stop playing. Speech is generated in chunks and queued. When the caller cuts in, the queue must be emptied and the sound already on the line must stop. If the agent finishes the current chunk, the caller hears a half second of the agent talking over them.
  2. Drop the unsaid text. The language model may have written a full paragraph, but the caller only heard the first two sentences. The record of the conversation should hold what was actually spoken. If it holds the whole paragraph, the agent may later say "as I told you, the late fee is Rs 300" when the caller never heard it.
  3. Keep the context. The caller's interruption is part of the same conversation. The agent should answer "ruko, woh date galat hai" with the loan, the date and the amount still in mind, not start again from the greeting.

Why quick stopping matters comes from how people talk. Stivers and colleagues compared ten languages and found the average gap between turns varied by about 250 milliseconds around a shared mean, with every language showing the same avoidance of overlap and short silences (Stivers et al., PNAS 2009, read 16 November 2026). A later review puts the gaps between turns at the order of 200 milliseconds (Levinson and Torreira, 2015, read 16 November 2026). Neither study tested an AI agent, and neither included an Indian language, so treat them as background, not a target. The lesson for barge-in is that people notice overlap fast. An agent that talks on for two seconds after "ruko" sounds like it is not listening.

What is the difference between a backchannel and a real interruption?

A backchannel is a short sound that says "I am following you", and a real interruption is the caller taking the floor. Both can arrive while the agent talks, and they must be treated differently.

In Indian calls the list is familiar. Backchannels: "haan", "hmm", "achha", "theek hai", "okay", "ji". Real interruptions: "ruko", "ek minute", "nahi nahi" and "wait". Your callers will have their own words, so test your own list. The same word can switch role with the situation. "Theek hai" while the agent explains is a backchannel. "Theek hai" after the agent asks "Shall I book Thursday?" is an answer, and the agent should have been waiting for it.

The decision after a sound is heard while the agent talks: noise, a backchannel and a real interruption each lead to a different action.
  • Caller says: "Haan, haan" while the agent reads a date

    What it is
    Backchannel
    What the agent should do
    Keep talking without a pause
  • Caller says: "Ruko, ruko" in the middle of an amount

    What it is
    Stop word
    What the agent should do
    Stop at once and listen
  • Caller says: "Nahi nahi, woh nahi, Thursday"

    What it is
    Correction
    What the agent should do
    Stop, drop the rest, answer the correction
  • Caller says: A cough or a horn

    What it is
    Noise
    What the agent should do
    Keep talking
  • Caller says: "Hello? Hello?" on a bad line

    What it is
    Real attempt to be heard
    What the agent should do
    Stop and check the caller is there

Our guide to Hinglish voice agents goes deeper on how these words mix across Hindi and English. The word lists are different in each language, which is why one global list does not work. See the languages we cover on our languages page.

Which steps should allow barge-in and which should not?

Allow barge-in on most steps, and restrict it on steps where the caller must hear every word. A single setting for the whole call cannot do both, so set it step by step.

The VoiceXML specification already says so: if the user must hear all of a warning, legal notice or advertisement, bargein should be disabled (W3C VoiceXML 2.0, section 4.1.5, read 16 November 2026). A voice agent has the same kind of steps.

  • Step: A required notice, such as a recording notice

    Barge-in
    Off or strict
    Why
    The caller must hear it in full
  • Step: A payment term or fee, read out

    Barge-in
    Off or strict
    Why
    A cut-off figure is a dispute later
  • Step: An OTP or reference number read back

    Barge-in
    Off or strict
    Why
    A half-heard number is a wrong number
  • Step: The first one or two seconds of the greeting

    Barge-in
    Strict
    Why
    A caller who says "hello?" is not interrupting
  • Step: Questions the agent asks

    Barge-in
    On
    Why
    A caller who knows the answer will give it early
  • Step: Explanations and long answers

    Barge-in
    On
    Why
    Callers should be able to say "ruko"

One more point. Restricting barge-in has a cost. A caller who cannot interrupt a long notice gets angry. Keep restricted steps short, and put them where the caller expects them.

Why does the agent stop when nobody spoke, or keep talking when somebody did?

Because the sound check is wrong in one direction or the other. The first is a detector that is too eager. The second is a stop word list that is too short, or a minimum speech length that is too long.

  • What you hear: The agent stops when nobody spoke

    Likely cause
    Sound sensitivity too high: traffic, a TV or a fan counts as speech
    What to change
    Lower the sound sensitivity, then retest on a noisy line
  • What you hear: The agent stops for every "haan"

    Likely cause
    Acknowledgement words are treated as interruptions
    What to change
    Make sure acknowledgement words never stop the agent
  • What you hear: The agent keeps talking over "ruko"

    Likely cause
    Stop words missing, or the minimum speech length is too long
    What to change
    Add stop words in the caller's language, then retest
  • What you hear: The agent hears itself and stops

    Likely cause
    Its own voice comes back through a speakerphone
    What to change
    Test on speakerphone, and lower sensitivity if needed
  • What you hear: The agent stops, then repeats the same sentence

    Likely cause
    The unsaid text was kept in memory, or the stop was a false alarm
    What to change
    Check what the transcript recorded as spoken
  • What you hear: A second voice in the room stops the agent

    Likely cause
    A family member or a TV is being counted as the caller
    What to change
    Lower sensitivity and keep the step short

Indian homes and streets give barge-in plenty to trip over: two-wheelers, a pressure cooker, a TV serial, a relative talking in the next room. Our post on Silero VAD on 8 kHz phone audio shows how those sounds look to a detector and how to tune for them.

How do you test barge-in before real calls?

Write a list of cut-ins, run each one by voice, and write down what the agent did. A text test cannot check barge-in, because the problem is in the audio.

A starter list:

  1. A clear "ruko" in the middle of a sentence. The agent should stop within a moment.
  2. A "haan" in the middle of a sentence. The agent should keep going.
  3. A cut-in in the first second of the greeting.
  4. A cut-in during a step where barge-in is off. The agent should finish the notice.
  5. A cough, a door closing and a horn. The agent should not stop.
  6. A call on speakerphone, so the agent's own voice is in the caller's line.
  7. A cut-in with a correction, such as "nahi nahi, Thursday". The agent should answer Thursday, not repeat itself.
  8. A long cut-in, where the caller talks for ten seconds. The agent should wait and not speak over them.

Run each one at least three times with different voices: a fast speaker, a soft one, and one with an accent your callers have. Change one setting at a time and rerun the failing case. Our guide on how to test a voice agent gives the wider method, and our post on endpointing and turn detection covers the other half of turn-taking.

How Bananaflow fits

Bananaflow is an AI voice agent platform for Indian businesses. Barge-in is set for each step of the call, so a recording notice can be protected while the rest of the call stays open. Acknowledgement words never stop the agent, and stop words do, in each language, with starter lists for all 11 languages that you can edit. A sound sensitivity setting, low, medium or high, decides how readily sound on the line counts as the caller talking. When a caller cuts in, the agent stops, listens and answers what they said.

You can test any agent in the browser, by voice or text, before it takes real calls. After each call you get a transcript and a recording. The inbound calls and outbound calls pages show how this fits each type of call.

What should you remember about barge-in?

  • Barge-in is the caller cutting in while the agent talks. It is not the call-centre meaning of a supervisor joining a call.
  • The agent decides from sound, minimum speech length and words. Each trades speed against accuracy.
  • On a cut-in, stop the audio, drop the unspoken text, and keep the conversation so far.
  • Acknowledgement words like "haan" should not stop the agent. Stop words like "ruko" should.
  • Restrict barge-in on steps the caller must hear in full, and keep those steps short.
  • Test by voice, with noise, speakerphone and real accents, and change one setting at a time.

FAQ

Frequently asked questions

  • What is barge-in in a voice agent?

    Barge-in is when a caller starts speaking while the voice agent is still talking, and the agent stops to listen. It lets callers cut in to correct, answer early or say stop. In voice AI this is the meaning. It is not the call-centre term where a supervisor joins a live call, which our glossary covers under call barging.

  • How does a voice agent detect that a caller is cutting in?

    It listens to the caller's line while it speaks. Voice activity detection tells it that someone is making sound, a minimum speech length filters out clicks and coughs, and the words, once recognised, show whether the sound was a real interruption or a short "haan". Systems differ in how many of these checks they use before they stop talking.

  • What happens to the agent's speech when a caller barges in?

    The audio that is still playing is cut off, and the words the caller never heard should not count as said. A well-built agent keeps what the caller did hear, drops the rest, and answers the new sentence. If it keeps the unheard text in its memory, it may later act as if the caller already knew it.

  • What is the difference between a backchannel and an interruption?

    A backchannel is a short sound that tells the speaker you are following, like "haan", "hmm" or "achha". An interruption is the caller taking the floor, like "ruko" or "nahi nahi, woh nahi". On Bananaflow, acknowledgement words never stop the agent, and stop words do, with separate lists for each language.

  • When should barge-in be switched off?

    Switch it off, or make it stricter, on a step the caller must hear in full, such as a recording notice, a payment term or an OTP read-back. The W3C VoiceXML specification says the same: if the user must hear all of a warning or legal notice, bargein should be disabled. On Bananaflow, barge-in is set for each step of the call.

  • Why does my voice agent stop talking when nobody spoke?

    Usually the voice activity detector counted noise as speech: a horn, a television, a second person in the room, or the agent's own voice coming back through a speakerphone. Lower the sound sensitivity and test again on a noisy line. If the agent still stops, check whether acknowledgement words are being treated as stop words.

  • How do you test barge-in before real calls?

    Write a short script of cut-ins and run each one by voice: a clean "ruko" mid-sentence, a "haan" that should not stop the agent, a cough, traffic noise, a speakerphone call and a cut-in during a step where barge-in is off. Record how the agent behaved each time. On Bananaflow you can test an agent in the browser by voice or text before it takes calls.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot