How it works
- Drop an audio or video file.
- Set the strength and output format.
- Download the cleaned file.
Why use it
- RNNoise neural denoiser in WebAssembly.
- Adjustable strength, so voices sound natural.
- Keeps the video stream untouched for video files.
- WAV, MP3 or M4A output with optional volume leveling.
Removing background noise from speech: what to expect
What it removes well, and what it does not
The noise remover uses RNNoise, a compact neural network built at Mozilla and Xiph.org to separate a speaking voice from the noise around it. It listens to the sound in slices of 10 milliseconds and decides, for each band of frequencies, how much belongs to the voice. We tested it on the same voice recording with different noises mixed in:
| Noise | Pauses got quieter by | Voice score before | Voice score after |
|---|---|---|---|
| Hiss, as loud as the voice | 23 dB | 0 dB | 13.1 dB |
| Hiss, 10 dB below the voice | 35 dB | 10 dB | 16.2 dB |
| Mains hum | 25 dB | 10 dB | 14.6 dB |
| Fan | 37 dB | 10 dB | 13.4 dB |
| Keyboard clicks | 6 dB | 10 dB | 14.1 dB |
| Music in the background | 7 dB | 10 dB | 12.0 dB |
Steady sounds, hiss, hum, fans, air conditioning, traffic and room rumble, are what it handles best. Sudden or changing sounds, such as clicks, door slams, dogs, other voices and music, are partly reduced at most. It is made for speech: on singing or instruments it treats much of the music as noise. The full test is in removing background noise from audio.
Choosing the strength
Full strength removes the most noise and suits noisy recordings. On recordings with only a little noise, full strength can thin the voice slightly, and a lower setting sounds more natural: in our test, about 50 percent gave the best result on light hiss. Even out the volume, on by default, raises the loudest moment to just below full scale after cleaning, since removing noise often leaves a recording quieter overall.
Videos keep their picture
Drop a video and, with Keep the video on, the cleaned sound is put back into it with the picture copied untouched: no loss of image quality, and the result plays like the original. Switch it off to save only the cleaned sound. For audio, WAV keeps the cleaned result exactly; MP3 and M4A give smaller files at 128 to 256 kbps.
Getting a better recording in the first place
- Get the microphone close to the mouth. Every halving of the distance makes the voice louder against the room.
- Turn off what you can: fans, air conditioning, notifications, the fridge.
- Record in a soft room: curtains, carpet and furniture absorb echo, which no noise remover can take out well.
If something goes wrong
- The voice sounds watery or robotic. The noise is very loud, or the strength is too high for light noise. Lower the strength.
- Music or clicks are still there. They are not steady noise. Cut those parts with the audio cutter instead.
- It says the file has no audio track. The video has no sound to clean.
How we tested. One CC0 voice recording from Mozilla Common Voice mixed with each noise at the levels shown, cleaned with this tool at full strength in Chrome. Levels of the pauses measured in dBFS before and after.
Questions and answers
What kinds of noise does it remove?
Steady and background noise such as hiss, hum, fans and traffic. It is designed for voices and can affect music.
Does it degrade the voice?
Full strength can sound slightly processed. Lower the strength to keep more of the natural room sound.
Is my recording uploaded?
No, everything runs on your device.
Related guides
Removing background noise from audio: what works
We mixed five kinds of noise into real recordings of people speaking and cleaned them with a small neural network that runs in the browser. Steady noise almost disappeared. Sudden sounds and music were another story.