Echo cancellation
In one sentence
Echo cancellation stops a device from hearing its own voice, so that a voice agent speaking through a laptop speaker does not pick up its own audio through the microphone and mistake it for the visitor interrupting.
Definition
Echo cancellation stops a device from hearing its own voice. Without it, a voice agent talking through a laptop speaker picks up its own audio through the microphone and decides it is being interrupted.
It is what lets a visitor talk over the agent on ordinary speakers, rather than only through headphones.
Acoustic echo cancellation is the most under-appreciated component in browser-based voice. Barge-in depends on it completely, and it fails quietly when it is misconfigured, so the symptom shows up somewhere else entirely.
The problem it solves
- The agent speaks through the speakers. That audio travels through the air and reaches the microphone.
- Without cancellation, the system hears speech, decides the visitor is talking, and cuts itself off.
- The visible symptom is an agent that interrupts itself within a second of starting to speak. It looks like a turn-taking bug, and it is not.
How it works
- The system already knows what it sent to the speakers. That is the reference signal.
- An adaptive filter models how that signal is changed by the room and the hardware before it reaches the microphone.
- The modeled echo is subtracted from the microphone input, leaving the visitor's speech behind.
- The filter keeps adapting, because moving a laptop or changing the volume changes the acoustic path.
Why it fails in practice
- Raw microphone capture. If the audio pipeline asks for unprocessed microphone input, the browser's built-in echo cancellation is bypassed. This is the most common cause, and it is a single configuration line.
- Non-linear distortion. Cheap speakers driven loudly distort, and distortion is harder to model and subtract.
- Long acoustic delay. Large or reverberant rooms stretch the echo path beyond what the filter models.
- Headphones make the problem disappear, which is exactly why it survives testing and then shows up in production.
Where it comes from
- WebRTC provides acoustic echo cancellation as a browser-native capability, alongside noise suppression and automatic gain control.
- That is one of the strongest practical arguments for using the browser's processed audio stream rather than raw capture.
Common misconception
That barge-in is a feature you switch on. Barge-in is a chain, and echo cancellation is the link that usually breaks. A platform that claims barge-in support without effective echo cancellation supports barge-in with headphones only.
Why it matters commercially
This is what decides whether a voice agent works on the hardware most people actually use. A demo run on a headset proves very little about a laptop sitting on a desk in an open-plan office.
In voice specifically
Text has no version of this problem. Nothing is played out loud and captured back, so there is no acoustic loop to cancel. Echo cancellation exists only because a voice device is producing sound and listening for sound at the same moment, through hardware sitting inches apart.
Where AsqVox fits
Running in the browser makes this matter more than it would in telephony, because laptop speakers with an open microphone are the normal setup rather than the exception. Browser-native echo cancellation is the mechanism that applies, which is one reason on-site voice leans on the browser's processed audio stream rather than raw capture.
Visual
The agent that interrupts itself
Barge-in is a chain. Echo cancellation is the link that usually breaks.
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.
WebRTC provides acoustic echo cancellation, noise suppression and automatic gain control as browser-native audio processing. Requesting raw or unprocessed microphone input bypasses all three.
AEC, NS, AGC built inindustry rangeWebRTC audio processing specification, 2026 - A capability fact, not a measurement. It is why the single most common echo bug is a one-line request for raw audio that turns the browser's own cancellation off.
Barge-in should cut synthesis within roughly 60 milliseconds of a detected interruption to feel natural.
About 60 msindustry rangeReal-time voice engineering convention, 2026 - A production budgeting figure rather than a published research result. It is a target for how fast the agent stops talking, not a claim about how fast the interruption is detected.
Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, which confirms that detecting an interruption is not the constraint. The difficulty is verification: telling the visitor apart from the agent's own echo.
Under 1 ms per 30 ms chunkvendor claimSilero VAD, 2026 - Reported by the project. Cited to locate the real bottleneck, which is separating a genuine interruption from the echo of the agent's own voice, not the speed of detection.
There is no published benchmark for echo cancellation effectiveness across browsers, devices and acoustic conditions, which is precisely the variation that decides whether barge-in works in the field.
-no reliable figureThe one thing you most need to know, how well it holds up on the hardware your visitors actually have, is the thing nobody measures. Test on real laptops and real speakers, not on a headset.
Examples
In practice
A voice widget runs perfectly in development on headsets and then fails on laptops, cutting itself off over and over. The cause is upstream: the media constraints ask for raw audio to hand the recognizer a cleaner signal, which switches off the browser's echo cancellation. Turning processed audio back on fixes it at once. An attempt to improve recognition quality had quietly broken turn-taking.
The everyday version
Echo cancellation stops the system from hearing itself through the speakers. Without it, your agent starts talking, hears its own voice come back through the microphone, assumes the customer cut in, and stops. It looks like a bug in the conversation. It is really a bug in the audio.
Usage
Who says it
- Audio and real-time engineers, routinely.
- Almost nobody in a business conversation, even though its absence is what generates the support tickets.
Where it turns up
- On a spec sheet it sits next to audio processing, barge-in support, supported devices and browser requirements.
- A claim of barge-in support should prompt one question in return: how does it perform on open speakers, not just on headphones.
Common misuse
- Requesting raw microphone input to improve recognition, which disables echo cancellation as a side effect.
- Testing exclusively on headsets.
- Blaming self-interruption on the turn detection logic when the real cause is the audio configuration.
Questions people ask
Why does my voice agent interrupt itself?
It is hearing its own voice through the speakers and mistaking it for the visitor. Almost always echo cancellation is off, usually because the audio pipeline requested raw microphone input, which bypasses the browser's built-in cancellation. It looks like a turn-taking bug, but it is an audio configuration bug, and it is typically a single line to fix.
What is acoustic echo cancellation?
It stops a device from hearing its own output. The system knows what it sent to the speakers, so an adaptive filter models how that reference signal is transformed by the room and hardware before reaching the microphone, then subtracts the modeled echo from the microphone input, leaving only the visitor's speech. The filter keeps adapting as the acoustic path changes.
Why does barge-in work with headphones but not on speakers?
Headphones remove the acoustic loop, so there is nothing for the microphone to pick up and nothing to cancel. On open speakers, barge-in depends entirely on echo cancellation. A platform that claims barge-in support without effective cancellation only supports it on headphones, which is why a headset demo proves very little. Test on speakers.
Where does browser echo cancellation come from?
WebRTC provides acoustic echo cancellation, noise suppression and automatic gain control as browser-native audio processing. Requesting raw or unprocessed microphone input bypasses all three, which is the most common reason echo cancellation is silently off in a browser-based voice agent.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.