Interruption handling latency
In one sentence
Interruption handling latency is how quickly a voice agent stops its own speech once a person starts talking over it, targeted at roughly 60 milliseconds and judged about an order of magnitude more strictly than the delay before it answers.
Definition
Interruption handling latency is how fast an agent stops talking once someone starts talking over it.
It is the mirror image of response latency, and it is held to a much stricter standard, because a slow stop is the machine talking over a person rather than merely a pause.
Response latency is measured in hundreds of milliseconds and interruption latency in tens, because the two are perceived differently. A delayed reply is a pause. A delayed stop is socially aggressive. People expect an interruption to be honored almost instantly, because that is how human conversation works. Roughly 60 milliseconds is the working target for cutting synthesis after a detected interruption; beyond a couple of hundred milliseconds it reads as talking over.
The chain that must complete inside that budget
- Detect the incoming speech.
- Verify it is the visitor and not the agent's own echo.
- Halt synthesis.
- Flush the audio already buffered downstream, which is the step most often missed.
- Reset conversational state to record what was actually heard rather than what was planned.
Why buffered audio is the hidden problem
- Stopping generation does not stop playback of audio already sent to the browser or the carrier.
- A system that halts synthesis but does not flush keeps speaking for as long as the buffer holds, which can be several hundred milliseconds.
- The visitor experiences the agent ignoring the interruption, while the logs show it was handled correctly.
The state problem that follows
- If the agent planned a three-sentence answer and was cut off after one, it has to record that only one sentence was heard.
- A system that records the planned response rather than the delivered one later references information the visitor never received, which produces baffling conversations.
Where measurement usually fails
- Measuring from detection to synthesis halt understates the real figure.
- The meaningful measurement is from speech onset to audio actually ceasing at the listener's ear, which includes the buffer flush and playback.
Common misconception
That interruption handling is a single feature toggle. It is a chain with several steps, and it is usually measured at the wrong point, which is why a platform can report a fast number and still feel like it talks over people.
Why it matters commercially
Interruptibility is what distinguishes a conversation from a recorded message, and it is the fastest way for a visitor to discover this is not the phone menu they were braced for. A slow reply is forgiven; a machine that keeps talking over you is not.
In voice specifically
Interrupting a text interface costs nothing, because you simply type. Interrupting speech requires the machine to notice you, verify it is really you, stop itself and discard what it had planned, all while its own voice is still coming out of the speaker.
Where AsqVox fits
AsqVox runs in the browser as a website voice agent, which makes the verification step harder, because laptop speakers and open microphones mean the agent's own voice reaches the microphone routinely. Echo cancellation is the dependency that lets verification tell the visitor apart from the agent.
Visual
Two latencies, two standards
A slow reply is a pause. A slow stop is rudeness.
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.
Interruption handling 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 response-latency target and still feel rude if it talks through an interruption.
The average gap between turns in human conversation is around 200 milliseconds, and people overlap speech routinely rather than waiting for clean boundaries.
200 msindustry rangeWidely cited conversational turn-taking baseline, 2026 - The second half 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.
Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, which confirms detection is not the constraint. Verification, buffer flush and playback control are where the budget is consumed.
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: if the stop feels slow, the time is going somewhere after detection.
WebRTC provides acoustic echo cancellation as a browser-native capability, which is the mechanism that lets verification distinguish visitor speech from the agent's own output on speaker-mode devices.
Browser-native echo cancellationindustry rangeWebRTC, browser standard, 2026 - A standard capability rather than a measured figure. Without it, interruption handling holds up only for headphone users, because the agent hears itself and stops for its own voice.
No published benchmark for interruption handling latency exists across voice agent platforms, and there is no standard measurement methodology. Claims of superior interruption handling are unverifiable by third parties.
-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.
No published research establishes the perceptual threshold at which delayed interruption handling becomes objectionable. The couple-of-hundred-milliseconds figure is a working convention, not a measured cliff.
-no reliable figureSo the 60 millisecond target and the point where a late stop starts to grate are both engineering conventions, and should be quoted that way.
Examples
In practice
A platform reports interruption handling at 40 milliseconds, measured from voice activity detection firing to the synthesis call being canceled. Users still report the agent talking over them. Measuring from speech onset to audio ceasing at the client shows 380 milliseconds, because several hundred milliseconds of audio were already buffered downstream. Adding an explicit buffer flush brings the real figure into range.
The everyday version
Interruption handling is how fast the agent shuts up when your customer starts talking. People forgive a machine that takes a moment to answer. They do not forgive one that keeps talking over them, which is why this has to happen roughly ten times faster than everything else.
Usage
Who says it
- Voice engineers, in turn-taking and barge-in work.
- Vendors, in platform feature comparisons, usually without a measurement methodology attached.
Where it turns up
- In a spec sheet, next to barge-in support, interruption handling, echo cancellation and full-duplex capability.
- In platform feature comparisons, where a superior-interruption-handling claim rarely comes with a way to check it.
Common misuse
- Measuring from detection to synthesis halt rather than to audio ceasing at the listener's ear, which understates the real figure.
- Treating interruption handling as a toggle rather than a chain of steps.
- Claiming support with no echo cancellation behind it, which limits it to headphone users.
Questions people ask
What is interruption handling latency?
It is how quickly a voice agent stops its own speech once a person starts talking over it. It is judged much more strictly than response latency, in tens of milliseconds rather than hundreds, with roughly 60 milliseconds as the working target. Behind it is a chain of steps, not a single toggle: detect, verify, halt synthesis, flush buffered audio and reset state.
How fast should a voice agent stop when interrupted?
Synthesis should stop within roughly 60 milliseconds of detected speech to feel natural. Past a couple of hundred milliseconds it stops reading as an interruption being honored and starts reading as the agent talking over you. Detection itself is not the constraint; the time goes on verification, flushing buffered audio and stopping playback.
Why does my agent keep talking after I interrupt it?
Because stopping generation does not stop the audio already sent to the browser or the carrier. If the system halts synthesis but never flushes that buffer, it keeps speaking for as long as the buffer holds, which can be several hundred milliseconds. The fix is an explicit buffer flush, and the tell is that the logs show the interruption handled while the visitor still heard it talk over them.
How is interruption handling latency measured?
From speech onset to the audio actually ceasing at the listener's ear, which includes the buffer flush and playback. Measuring from detection to synthesis halt understates it, sometimes by hundreds of milliseconds. There is no standard cross-platform benchmark, so a vendor figure cannot be verified by a third party. The practical test is to interrupt the agent live, on a laptop, with the speakers on.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.