Barge-in on AI voice calls: letting callers interrupt without breaking the conversation
Barge-in is a voice agent's ability to stop speaking when the caller starts talking and respond to what they said. Good barge-in reacts to real interruptions within a moment but ignores short sounds like "haan", "ji" or "hmm" and background noise. Tune it with short agent turns, a noise setting that matches the caller's surroundings, and real test calls.
- Without barge-in, callers must wait for the agent to finish, which feels like an old IVR.
- Short listening sounds such as haan, ji, hmm and okay should not stop the agent.
- On noisy lines, background sound can trigger false interruptions, so the setting must match the place.
- Short agent replies of one or two sentences reduce the need to interrupt at all.
- Test barge-in on real phone calls from a shop, a street and a quiet room, not in a text preview.
What barge-in means
The term comes from IVR systems, where it meant you could press a key or speak before the menu finished playing. On an AI voice call it means the same thing in conversation: while the agent is talking, the caller can cut in, and the agent stops, listens and answers the new point.
People interrupt all the time on the phone. They have already heard enough ("haan haan, I know the address"), they want to correct something ("no, Thursday, not Tuesday"), or they have a more urgent question. An agent that ploughs on regardless makes callers repeat themselves, and many simply hang up.
Barge-in sits next to latency as the thing that most decides whether a call feels natural. Our note on voice latency covers the other half: how fast the agent starts after the caller stops.
The two ways barge-in fails
| Failure | What the caller hears | Usual cause |
|---|---|---|
| Missed interruption | The agent keeps talking over them | Interruptions switched off, or speech detection set too strict |
| False interruption | The agent stops mid-sentence for no reason, then starts again | Background noise, a TV, traffic, or a short "hmm" taken as a new turn |
Fixing one tends to worsen the other. Make the agent very sensitive and it stops every time a bus passes. Make it deaf and it talks over the customer. The right balance depends on where your callers are.
Why Indian calls are a special case
- Listening sounds. Many callers say "haan", "ji", "achha", "hmm" or "okay" while the other person is talking, just to show they are following. These are not interruptions, and an agent that stops for each one feels jumpy.
- Noisy surroundings. Shop counters, markets, two-wheelers, factory floors and family homes with the TV on are normal settings for a business call in India.
- Speakerphone and shared phones. Other voices in the room can sound like the caller.
- Language switching. A caller may interrupt in a different language from the one the agent is speaking. The agent should follow them, not restart in the old language. See Hinglish voice agents.
So the goal is not "stop whenever there is sound". It is "stop when the caller is clearly taking a turn".
Design choices that reduce interruptions
The best barge-in is the one you rarely need. A few script habits help:
- Keep turns short. One or two sentences, then a question. Long monologues invite interruptions.
- Put the key fact first. "Your order ships tomorrow" before the explanation, not after it.
- Give prices as one final amount. "₹1,850 including delivery" is easier than a breakdown.
- After an interruption, answer the new point. Do not repeat the whole previous sentence.
- Acknowledge briefly. A quick "sure" or "haan" before the answer tells the caller they were heard.
Our guide to writing voice agent scripts has more examples.
Model providers are also working on this. OpenAI's changelog for 6 July 2026 lists improved silence and noise handling and interruption behaviour in GPT-Realtime-2.1. Speech-to-speech models build turn-taking into the model itself, while classic pipelines handle it in the voice layer. Either way, the settings and the script still decide how it feels to your caller.
Barge-in settings on Convo360
Each Convo360 voice agent has a Call behaviour section built for real-world Indian lines:
- Let the caller interrupt is on by default. The caller can speak at any moment, and the agent stops and answers the new point.
- Listening sounds are ignored. A short "हाँ", "जी", "hmm" or "okay" while the agent is talking is taken as the caller following along, so the agent carries on.
- Background noise handling can be set to Quiet, Normal or Noisy. Choose Noisy for a shop counter, street or factory floor.
- Acknowledgement style sets how the agent reacts before answering: a short word, a full phrase, none, or your own line.
- Advanced lets you fine-tune raw mic sensitivity and the silence-before-reply timing, or reset to the preset.
For a very noisy line, set Noisy and turn interruptions off. The agent then ignores background sound and finishes what it is saying. The agent also waits about half a second of silence before replying, so it does not answer a half-finished sentence. Details are in the voice docs.
How to test barge-in properly
A text preview cannot show you barge-in. You need real calls on real phones.
- Call the agent from a quiet room and interrupt it three times: once to correct a date, once to ask a new question, once with only "haan".
- Repeat from a noisy place, such as a shop or a roadside.
- Try speakerphone with someone else talking nearby.
- Read the transcripts in Calls: did the agent stop when it should, and carry on when it should?
- Change one setting at a time and repeat.
Keep a short note of which setting worked for which kind of caller. A dental clinic's front desk and a hardware shop's delivery line rarely need the same values. If one agent serves both kinds of caller, start with Normal and adjust after a week of real calls.
Step by step
- Shorten the agent's turnsRewrite replies to one or two sentences with the key fact first, so callers have less reason to cut in.
- Pick the noise setting for your callersIn Call behaviour choose Quiet, Normal or Noisy based on where most callers will be.
- Keep interruptions on unless lines are very noisyLeave Let the caller interrupt on by default, and turn it off only for lines where noise keeps cutting the agent off.
- Test with real interruptionsPlace calls that correct a detail, ask a new question and say only haan, from both quiet and noisy places.
- Review transcripts and adjust one thingRead each call in Calls, change a single setting, and test again until the agent stops and continues correctly.
Example conversation
Set up a voice agent that listens as well as it talks
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
What is barge-in in a voice bot?
Barge-in is when the caller starts speaking while the voice bot is still talking, and the bot stops to listen and respond. It comes from IVR systems, where it let callers skip a menu, and is now essential for natural AI phone conversations.
Why does my AI voice agent stop talking for no reason?
Usually background noise or a short listening sound is being treated as an interruption. Set the noise handling to match the caller's surroundings, and on very noisy lines turn interruptions off so the agent finishes its sentence.
Why does my AI agent talk over the customer?
Interruptions may be switched off, or speech detection may be set too strict for the line. Turn on the option that lets the caller interrupt, check the mic sensitivity, and keep the agent's replies short.
Will saying haan or hmm stop the AI agent?
It should not. A good voice agent treats short listening sounds like haan, ji, hmm or okay as the caller following along. On Convo360 these are ignored while the agent is talking, and real interruptions still stop it.
Can I test barge-in without making a phone call?
Not reliably. Barge-in depends on real audio, phone networks and background noise, which a text preview cannot reproduce. Place a few real test calls from quiet and noisy places, then read the transcripts.
Official sources
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.