Skip to content

Remove Background Noise from Audio — AI Noise Remover

Clean fan hum, hiss, traffic and room noise out of voice recordings and video audio with the RNNoise AI model, right in your browser.

On your device

Found a problem, or something that could be better? We read every message and we fix things quickly.

Remove background noise from audio in your browser: drop in a voice recording, podcast, voice memo, lecture or the soundtrack of a video, and the open-source RNNoise neural network strips out the steady noise behind the voice — computer fans, air conditioning, electrical hum, tape and preamp hiss, distant traffic, keyboard clatter. You can switch between the original and the cleaned version at the same point in the recording, choose how strong the cleanup is, and download the result as WAV or MP3.

RNNoise was trained on speech, so that is what it is for: it keeps voices and removes what is not a voice. It is not a music restorer, and on a song it can mute or warble the instruments; the page tells you when it hears little speech in a file. The model runs as WebAssembly inside your browser (a one-time download of about 3 MB, then cached) and processes roughly ten times faster than real time on a typical laptop, so a 10-minute file takes about a minute. Your file is never uploaded. To trim the result, use the audio cutter; to record a new take, the voice recorder.

How to use

  1. Drop an audio or video file onto the box, or click to choose one. Videos are fine: only their sound is used.
  2. Click Remove background noise. The model loads once, then a progress bar shows how far along it is.
  3. Press Play and switch between Original and Cleaned to compare; click the waveform to jump to a noisy passage.
  4. If the voice sounds watery or the pauses unnaturally dead, lower the Strength slider until it sounds natural.
  5. Pick WAV or MP3 and click Download cleaned audio.

Frequently asked questions

How do I remove background noise from a voice recording for free?

Open the file on this page and click Remove background noise. The RNNoise model cleans it in your browser, with no account, watermark or length-based fee; the only limit is 15 minutes per file, set by browser memory. Compare the result with the original, adjust the strength, and download it as WAV or MP3.

What kinds of noise can it remove?

Steady or slowly changing noise behind a voice works best: fans, air conditioning, fridge and electrical hum, hiss, rain, road noise, crowd murmur and keyboard typing. Sudden loud sounds that overlap a word (a door slam, a dog bark) are reduced but rarely vanish, and echo from a bare room is not removed, because RNNoise suppresses noise rather than reverberation.

Can I remove noise from a video?

Yes, the tool reads the soundtrack of MP4, MOV, WebM and most other video files. It exports cleaned audio only (WAV or MP3), not a new video; put the cleaned track back in your video editor, where you can also line it up with the picture.

Does it work on music?

Not well, and it is not meant to. RNNoise was trained to separate speech from noise, so it treats most instruments as noise to remove. The page warns you when it detects little speech in a file. For splitting a song into vocals and backing track, use the vocal remover instead.

Is my audio uploaded anywhere?

No. Decoding, noise removal and MP3/WAV encoding all happen on your device. The only downloads are the tool itself and, for unusual file types the browser cannot open on its own, the ffmpeg decoder from the jsDelivr CDN; your file stays on your computer or phone.

Why is the cleaned file mono?

RNNoise processes one channel, and speech recordings carry the same voice in both channels anyway, so the channels are averaged and the result is mono at 48 kHz. That is the normal format for voice-overs and podcasts, and it halves the file size.

How RNNoise removes noise

RNNoise, by Jean-Marc Valin at xiph.org (the people behind the Opus codec), combines classic signal processing with a small recurrent neural network. Every 10 milliseconds it splits the sound into a few dozen frequency bands, spaced the way the ear hears pitch, and the network estimates for each band how much of the energy is voice. Bands that are mostly noise are turned down, bands that are mostly voice are kept, and a pitch filter cleans up between the harmonics of the voice.

Because the network only decides gains, not the sound itself, it cannot invent speech or change what was said. Its typical failure is the opposite: on heavy noise it can thin out quiet syllables or leave a faint 'underwater' colour. The strength slider exists for that case; mixing 10–30 % of the original back in usually sounds more natural than 100 %.

Getting the best result

• Clean the recording before you edit it, and export WAV if you will process it further; MP3 is for the final file.

• Do not boost the volume or apply compression first: that raises the noise along with the voice and makes it harder to separate.

• Listen to the pauses and the ends of words with the Original/Cleaned switch. If breaths and word endings disappear, lower the strength.

• Prevent noise you can: a microphone closer to the mouth and away from the fan beats any noise remover. The voice recorder's voice-processing option also reduces noise while recording.

Limits worth knowing

Files up to 15 minutes are accepted, because the original, the cleaned copy and the playback buffers are all held in memory. Longer recordings can be split with the audio cutter and cleaned part by part. Processing runs on one processor core in the background, so the page stays usable; a slow phone may take several minutes for a long file.

Room echo (reverberation), clipping distortion and music in the background are not what RNNoise removes. For a song, the vocal remover can separate the voice from the instruments instead.