Barge-in
In one sentence
Barge-in is a voice agent stopping its own speech the instant a person starts talking over it, which is what makes the exchange feel like a conversation rather than a recording you have to sit through.
Not to be confused with Turn detection, or Endpointing.
Definition
Barge-in is being able to interrupt. You start talking, and the agent stops.
Being interruptible is most of what separates a voice agent from a phone menu, and people notice its absence immediately even when they cannot name it.
Five things have to land fast, and in sequence. While its own audio is still playing, the system needs to register that somebody else has started talking. It stops synthesis and playback. It clears whatever audio has already gone downstream. It drops the rest of the answer it was going to give. Last, it rewrites the conversational state so the new input is logged as a fresh turn instead of an aside tucked inside the previous one.
Why it is technically awkward
- Echo. The agent's own voice reaches the microphone, especially through laptop speakers. Without effective echo cancellation the system hears itself, calls it an interruption, and cuts itself off over and over.
- Buffered audio. Anything already handed off to the browser or the carrier keeps playing after synthesis has stopped. Emptying that buffer is the difference between a stop that feels instant and one that trails away.
- State. A half-delivered answer has to be recorded as half-delivered. Otherwise the agent believes it said something the visitor never heard, and refers back to it later.
The timing requirement
- Cutting synthesis within roughly 60 milliseconds of detected speech is what makes the stop feel natural.
- Past a couple of hundred milliseconds it stops reading as an interruption being honored and starts reading as the agent talking over the visitor.
Three variants
- Hard barge-in. Any detected speech stops the agent at once. Responsive, and exposed to every false positive from noise and backchannel sounds.
- Soft or confirmed barge-in. Waits for a minimum duration or word count before it stops. Harder to fool with noise, and a little slower off the mark.
- Backchannel-aware barge-in. Separates "mm-hm", which should not stop the agent, from a real interruption, which should. This is the sophisticated version, and it makes a material difference on long answers.
What the interruption rate tells you
- Constant interruption means the answers are too long. Treat the rate as a reading rather than only a feature: when it climbs, the agent is over-explaining.
Common misconception
That it is a toggle you switch on. What it actually is, is a chain of components, and the weakest link decides the outcome. That link is usually echo cancellation, which explains why the feature holds up in a headset demo and comes apart the moment somebody tries it on a laptop.
Why it matters commercially
Everybody has sat through a phone system that would not let them speak, and the irritation from it is learned deep. Cutting in and being heard is among the quickest ways for a visitor to work out that this is not that. Very little engineering, carrying a great deal of expectation-resetting.
In voice specifically
Interrupting a text interface costs nothing, because you simply type. Interrupting speech requires the machine to notice you, stop itself, and discard what it had planned to say, all while its own voice is still coming out of the speaker.
Where AsqVox fits
Running in a browser makes echo cancellation harder than it is on a phone line, because laptop speakers and open microphones are the norm rather than the exception.
Visual
The chain that has to fire in 60 milliseconds
Interruptible is the difference between a conversation and a menu.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Barge-in should cut synthesis within roughly 60 milliseconds of a detected interruption to feel natural.
60 msindustry rangeIndustry working threshold, 2026 - A working convention rather than a published standard, and a separate budget from time to first audio. A system can hit its latency target and still feel rude if it talks through an interruption.
Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, so noticing the interruption is not what consumes the budget. Verifying it, clearing buffered audio and stopping playback are.
Under 1 ms per 30 ms chunkvendor claimSilero VAD published benchmark, 2026 - Published by the project itself. It is quoted here to locate the bottleneck rather than to rate the detector: if barge-in feels slow, the time is going somewhere after detection.
The average gap between turns in human conversation is around 200 milliseconds, and people talk over each other constantly rather than waiting for tidy boundaries. Enforcing strict turns is a machine convention, not something people actually do.
200 msindustry rangeWidely cited conversational turn-taking baseline, 2026 - The second half of that claim is the part worth keeping. A system that refuses overlap is not being polite, it is being unlike every conversation its user has ever had.
No published benchmark exists for barge-in latency or accuracy across platforms, and there is no standard method for measuring it. Claims of superior interruption handling cannot currently be verified by a third party.
-no reliable figureWorth stating. It also points at the only test available to a buyer, which is to interrupt the agent in a live demo, on a laptop, with the speakers on.
Examples
In practice
A widget behaves perfectly in testing on headsets and falls apart on laptops, cutting itself off within a second of starting to speak. The cause is that acoustic echo cancellation was never engaged: the audio pipeline had been wired to raw microphone capture rather than the processed stream the browser offers. Turning echo cancellation on in the media constraints resolves it. No model change and no tuning, one line of configuration.
The everyday version
Barge-in is being able to interrupt. Think of an automated phone line working through nine options while you sit there knowing you want the third, with no way to skip ahead. Barge-in is the absence of that.
Usage
Who says it
- Voice engineers and contact center technologists. The word is much older than modern voice AI: in traditional IVR it meant a caller pressing a key or talking across a recorded prompt.
- Sales engineers, as a deliberate move in a demo, because interrupting on purpose is a strong live proof point.
Where it turns up
- In an RFP, sitting alongside echo cancellation, full-duplex support, conversational latency metrics and interruption handling.
- In legacy IVR requirement language, which often still specifies barge-in in its original keypad sense. Applied to a modern system that wording causes real confusion about what is being asked for.
Common misuse
- Advertising barge-in support with no echo cancellation behind it, which means the feature only holds up on headsets. The gap is real and it is common.
- Treating every detected sound as an interruption, so a visitor saying "right, yes" derails the answer.
- Using barge-in and turn detection interchangeably. Barge-in is the agent's turn being cut short. Turn detection is knowing when the visitor's turn ended.
Questions people ask
What is barge-in in a voice agent?
It is the agent stopping its own speech the moment the person starts talking. Behind that are five steps: detect the incoming speech, verify it is the visitor and not the agent's own echo, halt synthesis, flush the audio already sent downstream, and record what was actually heard rather than what was planned. It is a chain, not a switch.
Why does barge-in work with a headset but not on a laptop?
Because of echo. On a laptop the agent's own voice comes out of the speakers and back into the microphone, and without effective echo cancellation the system reads that as an interruption and cuts itself off. In one common version of the fault the audio pipeline was reading raw microphone capture instead of the processed stream the browser provides. Enabling echo cancellation in the media constraints fixes it, with no model change.
How fast does barge-in need to be?
Synthesis should stop within roughly 60 milliseconds of detected speech to feel natural. Past a couple of hundred milliseconds it reads as the agent talking over you. Detection itself is not the constraint, since Silero VAD scores a 30 millisecond chunk in under 1 millisecond. The time goes on verification, flushing buffered audio and stopping playback.
What does a high barge-in rate mean?
Usually that the answers are too long. Interruption rate is a diagnostic signal, not only a feature: if visitors cut in constantly, the agent is over-explaining and the fix is shorter responses rather than different interruption settings.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.