Free tool · AI noise model · Runs in your browser
Remove background noise from audio
Short answer
The tool runs the RNNoise speech model inside your browser, so nothing is uploaded. On 30 seconds of public domain speech with a real fan recording and a real street recording added at 5 dB below the voice, the Gentle setting raised the signal-to-distortion ratio by +4.8 dB (fan) and +4.7 dB (street) and made the pauses 8.9 and 9.7 dB quieter. FFmpeg's default denoiser changed the same files by 0.00 to +0.23 dB.
Nothing is uploaded: processing runs on your device. Maximum file size 100 MB. The first run loads a ~30 MB engine and a 298 KB noise model, once per visit.
Key takeaways
- Two settings: Gentle (mix 0.7, the default) and Standard (the model at full strength). Both use the RNNoise model "somnolent-hogwash", 297,646 bytes, trained for speech with steady background noise such as fans, air conditioning and computers.
- Gentle changed the SDR by +0.8 to +6.0 dB across eight test files (two noises at four levels). Standard changed it by -0.8 to +6.7 dB but made the pauses 15.3 to 26.1 dB quieter.
- The speech level moved by -2.7 to -0.2 dB with Gentle and -3.7 to -0.3 dB with Standard. The noisier the file, the more speech is lost along with the noise.
- FFmpeg's afftdn with its default settings did almost nothing (0.00 to +0.23 dB). With aggressive settings it handled the steady fan but not the street: +0.1 to +0.2 dB at the two loudest street levels, -3.4 to -0.8 dB at the two quietest.
- The test file was processed in under a second, and the output was a 48 kHz mono file. Files up to 100 MB are accepted.
How the cleanup works
FFmpeg describes its arnndn filter as one that reduces noise from speech using recurrent neural networks, with a model file that is always required and a mix option from -1 to 1 that sets how much of the filtered signal goes into the output (default 1). The tool loads the 297,646 byte model "somnolent-hogwash" in your browser and runs that filter on your device.
The model's own notes say it is meant for speech in a reasonable recording environment with fans, air conditioning and computers, and that "speech" means speech, not laughter or coughing. The author of the model collection states that, apart from its tools and README, none of the work is subject to copyright.
Standard runs the model at full strength. Gentle mixes in 70% of the filtered signal and 30% of the original, which keeps more of the voice and removes less of the noise.
Measured: speech with a real fan and a real street
Speech: 30 seconds of a LibriVox reading of The Heart of a Mystery (public domain). Noise: a studio fan recording (Hippocampuz, CC BY-SA 4.0) and a residential street recording (WrS.tm.pl, public domain), each added digitally at 0, 5, 10 and 15 dB below the speech level. The filters ran on the same single-thread WebAssembly engine your browser downloads.
Quality is reported two ways. SDR compares the cleaned file with the clean speech sample by sample, after finding the filter's delay (10 ms for the model, 25 ms for afftdn). The level figures use 20 ms frames: how many dB quieter the filter made the pauses, and how many dB the speech level moved.
| Noise and method | Noise added at 0 dB | at 5 dB | at 10 dB | at 15 dB |
|---|---|---|---|---|
| studio fan, AI model, Gentle | 5.9 dB (+6.0) | 9.8 dB (+4.8) | 13.1 dB (+3.1) | 15.8 dB (+0.8) |
| studio fan, AI model, Standard | 6.7 dB (+6.7) | 9.9 dB (+4.9) | 12.3 dB (+2.4) | 14.2 dB (-0.8) |
| studio fan, FFmpeg afftdn, strong (reduction 97 dB, floor -20 dB) | 4.8 dB (+4.9) | 10.5 dB (+5.5) | 12.1 dB (+2.1) | 12.4 dB (-2.6) |
| studio fan, FFmpeg afftdn, defaults (reduction 12 dB, floor -50 dB) | 0.0 dB (0.0) | 5.0 dB (0.0) | 10.1 dB (+0.1) | 15.2 dB (+0.2) |
| residential street, AI model, Gentle | 5.5 dB (+5.6) | 9.7 dB (+4.7) | 13.5 dB (+3.5) | 16.6 dB (+1.6) |
| residential street, AI model, Standard | 6.7 dB (+6.7) | 10.1 dB (+5.1) | 13.1 dB (+3.1) | 15.1 dB (+0.1) |
| residential street, FFmpeg afftdn, strong (reduction 97 dB, floor -20 dB) | 0.2 dB (+0.2) | 5.0 dB (+0.1) | 9.2 dB (-0.8) | 11.6 dB (-3.4) |
| residential street, FFmpeg afftdn, defaults (reduction 12 dB, floor -50 dB) | 0.0 dB (0.0) | 5.0 dB (0.0) | 10.0 dB (0.0) | 15.0 dB (0.0) |
One narrator, two noise recordings and noise added digitally, not recorded in the same room as the voice. We did not test music, several speakers, reverberation or noise louder than the voice.
What the pauses and the speech did
SDR rewards getting close to the clean waveform. It cannot tell you how the pauses sound, so the next two tables use level instead. At 5 dB, Standard made the pauses 18.4 dB (fan) and 25.3 dB (street) quieter, Gentle 8.9 and 9.7 dB.
| Method | studio fan: pauses, speech | residential street: pauses, speech |
|---|---|---|
| AI model, Gentle | 8.9 dB quieter, -1.1 dB | 9.7 dB quieter, -1.0 dB |
| AI model, Standard | 18.4 dB quieter, -1.4 dB | 25.3 dB quieter, -1.3 dB |
| FFmpeg afftdn, strong (reduction 97 dB, floor -20 dB) | 13.7 dB quieter, -0.9 dB | 1.1 dB quieter, -0.5 dB |
| FFmpeg afftdn, defaults (reduction 12 dB, floor -50 dB) | 0.0 dB quieter, 0.0 dB | 0.0 dB quieter, 0.0 dB |
The same measurement at 15 dB
A quieter background gives the model less to remove and more chance of touching the voice. At 15 dB Standard still made the pauses 15.3 to 22.5 dB quieter, but its SDR change was -0.76 to +0.12 dB, which is why Gentle is the default.
| Method | studio fan: pauses, speech | residential street: pauses, speech |
|---|---|---|
| AI model, Gentle | 8.0 dB quieter, -0.3 dB | 9.3 dB quieter, -0.2 dB |
| AI model, Standard | 15.3 dB quieter, -0.3 dB | 22.5 dB quieter, -0.3 dB |
| FFmpeg afftdn, strong (reduction 97 dB, floor -20 dB) | 39.6 dB quieter, -0.6 dB | 6.5 dB quieter, -0.5 dB |
| FFmpeg afftdn, defaults (reduction 12 dB, floor -50 dB) | 0.4 dB quieter, 0.0 dB | 0.1 dB quieter, 0.0 dB |
Gentle or Standard
The measured pattern is simple. The noisier the recording, the more Standard helps; the cleaner the recording, the more it risks the voice.
| Noise added at | Gentle: SDR change | Standard: SDR change | Standard: pauses quieter by |
|---|---|---|---|
| 0 dB | +6.0 / +5.6 dB | +6.7 / +6.7 dB | 21.5 / 26.1 dB |
| 5 dB | +4.8 / +4.7 dB | +4.9 / +5.1 dB | 18.4 / 25.3 dB |
| 10 dB | +3.1 / +3.5 dB | +2.4 / +3.1 dB | 16.5 / 24.2 dB |
| 15 dB | +0.8 / +1.6 dB | -0.8 / +0.1 dB | 15.3 / 22.5 dB |
Use Gentle when the noise is a background you can still talk over. Use Standard when the noise is nearly as loud as the voice and you would rather lose some warmth than keep the noise. If the voice sounds weaker, go back to Gentle.
Why FFmpeg's defaults are not enough
FFmpeg's afftdn filter denoises with an FFT. With its defaults (reduction 12 dB, noise floor -50 dB) it changed our files by 0.00 to +0.23 dB, so it does almost nothing on these recordings. With the strongest settings (reduction 97 dB, noise floor -20 dB) it cleaned the fan, which is steady, by +2.1 to +5.5 dB at the three noisier levels, and it lowered the SDR by 2.6 dB at the quietest level. On the street recording the change was +0.1 to +0.2 dB at the two noisier levels and -3.4 to -0.8 dB at the two quieter ones.
The reason is how the filter works: it estimates a noise floor and subtracts it, which suits hiss and hum and fails on noise that changes, like traffic. A speech model separates voice from noise instead of subtracting a fixed floor.
What we did not test
- Music, singing, laughter and other sounds that are not speech. The model's notes say speech only.
- More than one speaker, or a speaker who moves away from the microphone.
- Echo and reverberation. The notes list fans, air conditioning and computers as the intended noise.
- Noise that is louder than the voice (below 0 dB).
- Whether the MP3 output sounds different from the WAV we measured: the tool encodes MP3 at 192 kbps, and we measured the WAV path.
Settings and limits
The tool runs an arnndn filter with the model loaded from this site, then encodes MP3 at 192 kbps or WAV at 16 bits. Files up to 100 MB are accepted. In our headless run of the same engine the 30 second files took 0.1 to 0.6 seconds; your device will differ.
For a video file, use the video version of this tool: it keeps the picture exactly as it is.
Free tools:Remove background noise from videoNormalize audio loudnessRemove silence from audioExtract MP3 from a video
Questions people ask
Does my audio get uploaded?
No. The tool runs ffmpeg and the noise model inside your browser. The first run downloads an engine of about 30 MB and a model of 298 KB, once per visit.
Which setting should I pick?
Gentle for most recordings. On our files it changed the SDR by +0.8 to +6.0 dB and kept the speech level within -2.7 to -0.2 dB. Standard removes more noise and lowers the speech level more.
Can it remove music or other voices?
It is built for speech in steady background noise. Music and other voices were not tested, and the model notes say it is for speech.
Why did the noise barely change with FFmpeg's default denoiser?
Its default settings are mild. With them, afftdn changed our test files by 0.00 to +0.23 dB. The tool uses a neural network model instead.
Does it work on traffic or crowd noise?
On a residential street recording the Gentle setting changed SDR by +1.6 to +5.6 dB across four levels. That is one recording, so try your own file.
What does the output look like?
An MP3 at 192 kbps or a 16-bit WAV. In our run the WAV was mono at 48 kHz, the rate the model works at.
How long does it take?
In a headless run of the same engine, 30 seconds of audio took 0.1 to 0.6 seconds.
Can I use it on a video?
Yes, with the video version. It copies the picture unchanged and re-encodes only the sound.
Is the noise model free to use?
Its author says that, apart from the training tools and the README, none of the model work is subject to copyright. The README is linked in the sources.
Related
Sources and how we checked
- FFmpeg Filters Documentation: arnndn, afftdn and asdr (read 2026-10-06)
- GregorR/rnnoise-models: README and model notes (somnolent-hogwash) (read 2026-10-09)
- Internet Archive: The Heart of a Mystery, LibriVox recording (public domain mark) (read 2026-10-09)
- Wikimedia Commons: Ventilador de estudio.ogg by Hippocampuz (CC BY-SA 4.0) (read 2026-10-09)
- Wikimedia Commons: 20220623 Noise pollution, residential area, by WrS.tm.pl (public domain) (read 2026-10-09)
Reviewed October 9, 2026. The measurements ran the tool’s own filter arguments on the same single-thread WebAssembly engine your browser downloads. The scripts, the results and the source recordings’ licenses are in data/seo/hub/evidence/2026-10-09/noise.