Noise suppression

Speech Technologyalso: noise reductionalso: noise cancellation

In one sentence

Noise suppression removes background sound from captured audio so that speech is clearer, and it is what lets a voice agent work for someone in a cafe rather than only in a quiet room.

2.7% clean, 8 to 12% real world Whisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026Last reviewed 31 July 2026

Definition

Noise suppression removes background sound from captured audio so that what is left is closer to just the speech. It is one reason a voice agent can work for a visitor who is not sitting in a quiet room.

It is not the same as echo cancellation. Echo cancellation removes a sound the system already knows about, its own voice. Suppression removes the sounds it does not.

Noise suppression and echo cancellation are frequently bundled together and solve different problems. Echo cancellation removes a known signal, the system's own output coming back through the microphone. Noise suppression removes unknown signals, meaning everything that is not the speaker.

The approaches, and what each is good for

  • Spectral subtraction and other statistical methods estimate the noise profile during the gaps between speech and subtract it. Effective on steady noise such as a fan or traffic hum, poor on sudden sounds.
  • Neural suppression uses models trained to separate speech from noise, and handles non-stationary noise such as a door slamming or a second conversation far better.
  • Browser-native suppression is provided as part of the WebRTC audio processing chain, alongside acoustic echo cancellation and automatic gain control.

The tradeoff that catches people out

  • Aggressive suppression removes noise and damages speech with it. Over-suppressed audio sounds processed and can lose the quiet consonants that carry meaning.
  • Recognition accuracy can fall as suppression gets more aggressive, because the model was trained on audio that includes some noise and the artifacts of suppression are unfamiliar to it.
  • So a setting can sound cleaner to a person and perform worse for the recognizer. That is counterintuitive, and it is worth testing rather than assuming.

Where it matters most

  • Mobile visitors, who are often in uncontrolled environments.
  • Open-plan offices, where the dominant noise is other people's speech, the hardest case because it is spectrally identical to the target.
  • Homes with a television on, which produces speech-like sound that can trigger voice activity detection.

The one thing it cannot fix

  • Another person talking nearby is not noise in the technical sense. It is speech that belongs to someone else.
  • Suppression struggles with this. Speaker separation or a directional microphone array handles it better, and neither is available in a typical browser.
  • This is a real limit of browser-based voice in shared spaces, and offering a text option is the honest answer.

Common misconception

That more noise suppression is always better. Beyond a point it degrades recognition, and the best setting for a machine transcribing the audio can differ from the best setting for a person listening to it.

Why it matters commercially

Noise conditions decide whether a voice agent works for the visitors who are not sitting quietly at a desk, which on mobile is most of them.

In voice specifically

A text chat has no microphone, so the room never enters it. In speech the background is part of the input, and whatever the recognizer gets wrong because of the room caps every number downstream of it.

Where AsqVox fits

Browser-native noise suppression is the mechanism the Orb relies on. Where suppression cannot help, in shared spaces with competing speech, the text chat fallback is the correct answer rather than a degraded call.

Visual

Two different problems, often bundled

Two different problems, often bundled60recognition accuracy, percent70recognition accuracy, percent80recognition accuracy, percent90recognition accuracy, percent100recognition accuracy, percent110recognition accuracy, percent84recognition accuracy, percentOff90recognition accuracy, percentLight93recognition accuracy, percentModerate88recognition accuracy, percentAggressive81recognition accuracy, percentMaxSource: Illustrative shape, not measured dataAccuracy peaks at Moderate. Perceived cleanliness keeps climbing past it.query types absorbed, widest and earliest firstSteady noise: a fan, traffic hum, an air conditionerSudden noise: a door slam or a siren, handled better by neural modelsA television in the room, speech-like but not the speakerA second person talking nearby, spectrally identical to your visitor

More suppression is not better. Test it against the recognizer, not your ears.

Two things are being traded here. Perceived cleanliness rises steadily as suppression gets more aggressive; recognition accuracy rises, peaks, then falls, because past a point the artifacts of suppression cost the recognizer more than the noise did. The two problems people bundle are different: echo cancellation removes a signal the system already knows, its own output, while noise suppression removes signals it has never seen. The bars below rank what suppression can handle, widest first, down to the one it cannot, a second person talking, which is not noise but speech belonging to someone else, where text chat is the honest fallback. The curve is an illustrative shape, not measured data, because no public benchmark maps accuracy against suppression aggressiveness.

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 noise suppression as part of its browser-native audio processing chain, alongside acoustic echo cancellation and automatic gain control.

Browser-nativeindustry range

WebRTC audio processing chain, 2026 - A capability of the browser stack rather than a measured number. Requesting raw or unprocessed microphone input bypasses it.

Neural suppression handles non-stationary noise such as a door slam or a second conversation substantially better than spectral subtraction, which performs adequately on steady noise only.

-industry range

Audio engineering convention, 2026 - A qualitative distinction between two families of method, not a benchmarked gap. Steady noise means fans and traffic hum; non-stationary means sudden or changing sound.

Whisper Large-v3 records around 2.7 percent word error rate on clean benchmark audio and roughly 8 to 12 percent on real-world English, and noise is a primary contributor to that gap.

2.7% clean, 8 to 12% real worldindependent

Whisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026 - The benchmark half is a published result on a public test set; the real-world half is a reported band, so quote it as a range. Noise is one of the conditions that widens the gap.

Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate under production conditions.

5.26% to 6.84%vendor claim

Deepgram Nova-3 reported figures, 2026 - Published by the provider. A production-conditions figure, which is more useful than a clean benchmark and still not comparable across vendors unless the test set and normalization match.

No published benchmark measures recognition accuracy as a function of suppression aggressiveness, which is exactly the curve a deployment would need to tune against.

-no reliable figure

State the absence. The direction is known, that too much suppression eventually hurts, but the operating point has to be found by testing on your own audio.

No standard benchmark measures noise suppression performance across realistic consumer conditions.

-no reliable figure

Vendors describe their approaches; independent comparison across the conditions real visitors are in does not exist publicly.

Examples

In practice

A team turns noise suppression up to help mobile visitors and watches word error rate climb instead of fall. The aggressive setting is stripping out quiet fricatives along with the noise, and the recognizer turns out to have been more tolerant of the original noise than of the suppression artifacts. Dialing the aggressiveness back down improves accuracy, at the cost of audio that sounds slightly less clean to a human ear.

The everyday version

Noise suppression cleans up background sound so the system can hear your customer. It does well on steady noise like traffic or a fan. It does badly when the background is another person talking, because that sounds exactly like the thing it is trying to keep. That is why a text option matters for people in busy places.

Usage

Who says it

  • Audio engineers and real-time communication teams use the term routinely.
  • It is rarely a business word, though its effects show up in support complaints about recognition failing in noisy places.

Where it turns up

  • On a spec sheet it sits next to audio processing, supported environments, mobile performance and accessibility.

Common misuse

  • Turning suppression up on the assumption that cleaner audio is always better for recognition.
  • Expecting it to handle competing speech, which it cannot.
  • Treating it as the same thing as echo cancellation, which solves a different problem.

Questions people ask

What is the difference between noise suppression and echo cancellation?

They are bundled together and solve different problems. Echo cancellation removes a known signal, the system's own output coming back through the microphone. Noise suppression removes unknown signals, meaning everything that is not the speaker. A platform can have one working and the other misconfigured.

Does more noise suppression improve speech recognition?

Only up to a point. Recognition accuracy rises then falls as suppression gets more aggressive, because the model was trained on audio that includes some noise and the artifacts of over-suppression are unfamiliar to it. A setting can sound cleaner to a person and perform worse for the recognizer, so test it against the recognizer rather than your ears.

Can noise suppression remove someone talking in the background?

Poorly. A second person nearby is not noise in the technical sense, it is speech that belongs to someone else, and it is spectrally identical to your visitor. Speaker separation or a directional microphone array handles it better, and neither is available in a typical browser, so a text fallback is the honest answer in shared spaces.

Where does noise suppression matter most?

For visitors who are not at a quiet desk: mobile users in uncontrolled environments, open-plan offices where the dominant noise is other people talking, and homes with a television producing speech-like sound. On mobile that is most of the traffic, which is why noise handling decides whether a voice agent is usable at all.

Share this definition

Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.