Remove Background Noise from Audio

Steady, unchanging background: a fan, air conditioning, tape hiss, mains hum, the drone of traffic through a window, a computer in the corner, keyboard clatter under speech. Those are the sounds the model can learn the shape of and subtract. It is much weaker on one-off events — a door slamming, a cough, a chair scraping — and it cannot separate another person talking in the background from the person you want, because to the model both are speech.

Yes. No signup, no credit card, no daily minute allowance and no watermark or voice-over on the result. The model runs on your own computer, so there is no server time for us to bill you for. You can clean ten files in a row if you want to.

No. The file is read straight from your disk, decoded by the browser and processed by a WebAssembly build of the noise model inside the tab. You can check it yourself: open DevTools, switch to the Network panel and clean something — you will see the model being downloaded the first time, and your recording going nowhere.

No, and it is worth being blunt about it: the model was trained on speech, so anything that is not a voice looks like noise to it. Point it at a song and it will thin out the cymbals, hollow out the reverb and chew on sustained notes. Use it on interviews, voice memos, lectures, podcasts and voice-overs — not on a track you are mixing. If you only want a little of the room gone from a recording that also has music in it, drop the Reduction slider to 40 or 50 and listen carefully.

It is a blend between your original and the cleaned version, not a second pass through the model — which is why moving it is instant. At 100% you get everything the model is willing to remove; some voices start to sound thin or slightly underwater there. The "keep a little room tone" preset sits at 85%, which leaves a thin layer of the original room in: total silence between sentences sounds unnatural, and a faint room tone also hides the moments where the model switches on and off. There is also a gentle 80 Hz filter for desk thumps and rumble, off by default.

MP3, WAV, M4A, AAC, OGG, Opus and FLAC, plus MP4, MOV and WebM video. A video file is accepted on purpose — a phone recording of someone talking is one of the most common things people need cleaned — but only the audio comes back out. You get an audio file, not a video with a new soundtrack.

Up to 200 MB or 60 minutes. The real constraint is memory, not the file size: everything is decoded to 48 kHz floating-point audio, which costs about 11 MB per minute per channel, and the tool holds the original, the cleaned copy and the mix at the same time. A ten-minute interview is comfortable on a laptop; past about fifteen minutes you will see a warning, and it is worth closing other tabs. Processing itself is a good deal faster than listening to the recording would be.

Yes, in any modern mobile browser — the model is small and downloads once, around a megabyte. Keep the files shorter than on a laptop, though: a phone has far less memory to play with, and a long stereo recording can exhaust the tab. If you also want an MP3 rather than the WAV, that step pulls in a one-off 31 MB audio engine; the WAV download needs nothing extra and is ready the moment the cleaning finishes.