Background noise remover test: 4 real-world noises

How we tested

Each sample is 12 seconds of a narrator from a public-domain LibriVox recording, mixed with a real noise recording released under CC0 on Freesound: rush-hour city traffic, an air-conditioning vent, a busy café with a milk frother, and strong outdoor wind. Speech and noise were each normalised to a similar loudness before mixing, so the noise sits only a few decibels below the voice. That is harsher than most real recordings.

Every mix went through both of Denoisa’s modes. Standard runs DeepFilterNet3, an open-source speech enhancement model, directly in the browser; the clips below come from that exact model. Enhanced runs a larger speech model on our servers. Both received the same file, and the outputs are time-aligned with the original, so you can switch between them while a clip plays.

We measured three things on 50-millisecond slices of each file: the noise floor (the quietest 10% of slices, which is mostly the gaps between words), the speech level (the loudest 5%), and how much of the energy in the speech itself sits above 4 kHz. That last number is where consonants like s, sh, t and f live. Take them away and a voice sounds muffled, even if the noise is gone.

Steady hum: easy to remove, easy to overdo

An air conditioner produces almost the same sound every second, so both modes pushed it far down: Standard to −72 dB and Enhanced to −93 dB, which is effectively digital silence between words. But Standard also took some brightness out of the voice: the share of speech energy above 4 kHz fell from −22.7 dB to −35.4 dB, while Enhanced stayed close to the original at −24.2 dB. Hum sits in the same low and middle frequencies as a voice, and a small model with little context cuts a bit too much.

Traffic and wind: noise that moves

Traffic and wind change from moment to moment and cover a wide range of frequencies, including the ones a voice uses. A fixed noise profile, the approach older tools take, cannot follow them. Neural models handle this by deciding, every few milliseconds, how much of each frequency band belongs to the speaker.

On street traffic the two modes were practically tied: the noise floor dropped from −19 dB to −45 dB with both, and both kept the voice bright. Wind was different. Both removed it well (−56 dB with Standard, −58 dB with Enhanced), but Standard again dulled the voice: speech energy above 4 kHz went from −26.5 dB to −32.0 dB, while Enhanced brought it up to −16.0 dB. Wind rumble overlaps the voice heavily, and that is where a larger model pays off.

Café chatter: when the noise is also speech

Background conversation is often called the hardest noise to remove, because it is made of the same thing you want to keep. In this sample the two modes were close: the noise floor fell from −27 dB to −67 dB with Standard and −66 dB with Enhanced, and both kept most of the voice’s high-frequency detail (−18.7 dB and −14.6 dB, against −19.6 dB in the original). Enhanced is slightly brighter. Switch between the versions below and listen to the speaker’s s and t sounds.

Less noise is not the same as a clearer voice

Put side by side, the noise numbers alone say the two modes are similar. The difference shows up in the voice: on hum and wind, Standard removed almost as much noise as Enhanced but also part of the consonants, so the speech sounds duller. On traffic and chatter it held up.

This is the trade-off every noise remover makes. When the model is unsure whether a sound is noise or voice, it can either keep it, leaving some noise, or remove it, dulling the voice. A metric that only measures leftover noise rewards the second choice, and listeners notice.

Why a model in your browser makes this trade-off

Standard has to be small enough to download and fast enough to run on an ordinary laptop or phone. The DeepFilterNet3 model we use is 16 MB, and the first visit downloads about 30 MB in total, including the runtime. After that it cleans audio roughly five times faster than real time on a recent laptop: a 6-minute file took about 25 seconds on four threads. It processes sound in 10-millisecond steps and looks only 20 milliseconds ahead, so it decides with very little context.

In return your file never leaves your device. Nothing is uploaded and there is nothing for us to store. A larger model on a server can look at more context and spend more computation per second of audio, which is what lets it keep consonants that a small model gives up.

Which one to use

For traffic, general room noise and background chatter, Standard is usually enough, and files up to 5 minutes are free. For hum or wind, or whenever the cleaned voice sounds dull, try Enhanced. Every file gets a free 30-second preview of both, so you can compare on your own recording before using any credits.

What this test does not show

This is a small test: four 12-second samples with one narrator each. The noise was mixed in artificially, and the measurements compare each file with itself rather than with the clean original, so they describe noise and brightness, not overall quality. Real recordings also bring echo, cheap microphones and clipping, which we did not test. Treat the results as a guide to what kinds of noise are hard, not as a benchmark score.