Drop a recording, hear it with the fan, the hiss and the hum taken out, and download it if it sounds better. No watermark, no account, no upgrade prompt. Nothing is uploaded, so there is no cap to hit.
Four recordings — three real, one openly mixed — each put through this page and cut to about five seconds. Press play and use the switch. Nothing is downloaded until you do.
An amateur interview with a low electrical hum sitting under the whole recording. Steady noise behind a voice is what this tool is built for.
Listen to the gaps between sentences. The hum is gone from them and the voice is left where it was.
Between words, 21.8 dB out. On the words, 0.6 dB — nothing you can hear.
Margarita Monet interview, archive.org, CC0 1.0
A street parade recorded in 1982: wind across the microphone and the general roar of people standing around.
A crowd heard as one blur is steady noise, and it comes out much like a fan does. Individual people talking right behind you is a different thing, and is below.
Between words, 14.6 dB out. On the words, 1.0 dB.
Santa Ynez Centennial Parade 1982, Santa Ynez Valley Historical Museum / California Revealed, archive.org, Public Domain Mark 1.0
An oral history taped in 1980. The hiss is loud enough that the voice is only about 11 dB above it.
The pauses go quiet and the voice is untouched — and it is still hard to listen to. The hiss sitting on the words is not something this can reach. It takes noise out from around a voice, never from under one, and on a recording this noisy that leaves most of the problem where it was.
Between words, 12.2 dB out. On the words, 0.6 dB.
Palm Desert Oral History: Dan Callahan 1980, Historical Society of Palm Desert / California Revealed, archive.org, Public Domain Mark 1.0
A public-domain poetry reading with real concert applause mixed underneath it — a made file, mixed at a known level, because no freely licensed recording of speech with applause exists to borrow.
The applause between the phrases drops away and the voice comes through clearly where it used to sit in the noise. But listen to the pauses: the loudest single claps still poke through almost untouched. Steady crowd wash reads as noise to this tool; one sharp clap looks too much like a consonant for it to touch.
Between the phrases, 15.9 dB of the applause is gone. On the words, the voice moved 1.8 dB. The loudest clap, 2.0 dB — you still hear it.
Mixed from The Rime of the Ancient Mariner, LibriVox, PD Mark 1.0, and Sound Effects — Applause after a concert, Amada44, Wikimedia Commons, CC0 1.0
Two things with no clip here. Background chatter — a café, an open-plan office, one or two people talking right behind you — is the case it cannot do, because that noise is speech and there is no reliable way to tell whose voice is whose. Measured on a recording mixed at a known noise level, it moved the voice against the noise from 5.00 dB to 5.58 dB, which is no change anyone would hear. And it is built for speech: hand it music and it will gut it. Measured on a clean guitar recording, it took 12 dB off everything between 2 and 6 kHz and 15 dB off everything above that. If there is a performance under the noise rather than a voice, this is the wrong page, whatever is on top of it.
Before and after, one switch, same place in the recording. If it does not sound better, do not download it.
No account, no trial, no upgrade prompt on the download. Ads pay for the site.
That is why there is no cap and no account. Turn your network off once it starts and it still finishes.
Same length, same channels, same loudness, same kind of file. Only the noise behind the voice is touched.
memo-clean-8f2a1c.mp3.It is built for a voice with something steady behind it: a computer fan, an air conditioner, tape hiss, a mains hum, traffic outside, the general hum of a room. That is the noise it strips out most cleanly, and the difference is usually obvious in the silences between sentences.
It is weakest at background chatter — a café, an open-plan office, other people talking behind you. That noise is speech, with the same shape as the voice you want to keep, and no tool can reliably tell one from the other. Expect little improvement there, and check by listening rather than by hoping.
Very old or very narrow recordings are the other hard case. A 1940s broadcast or a bad phone line carries so little of the voice that parts of it can be mistaken for noise and thinned out. That is what Gentle is for: it puts half of your original back, so less noise comes out but the speech is left much closer to how it started.
People arrive here wanting a clap, a burst of applause or a stray sound effect taken out, and those are exactly the sounds a steady-noise remover is worst placed to judge — so we measured it instead of guessing. We mixed synthetic claps, a run of applause, a two-tone effect and a robot-treated voice into a spoken sentence, and ran each file through the same filter and the same model this page uses.
Applause between sentences comes down hard. Our two-and-a-half-second burst lost 30 to 40 dB — foreground to barely there — and the tool cut deeper the longer the clapping ran. A lone clap is the opposite story: the first clap in our test came through 1.7 dB quieter, which your ear reads as not removed at all, and the second and third were only partly cut. It takes roughly half a second of continuous clapping before the model decides it is noise, so the front of every burst leaks through, and anything that claps once — a door, a dropped box, a single pair of hands — simply stays. So does any clap that lands while someone is talking, and that part is by design: the tool’s first duty is to keep whatever overlaps speech, because taking it out would take the words with it.
Sound effects survived our test almost untouched. A clean two-tone beep came out 2.8 dB quieter, which is audibly identical — a tone looks nothing like the hiss and hum the model was trained on. If what you need is an effect lifted out of a recording, this page is not that tool, and we would rather say so here than have you find out after the download.
And a robotic-sounding voice comes back robotic. We ran a deliberately robotised recording through and the output measured within a tenth of a decibel of the input: the model decided the robot voice was the voice, and protected it. That distortion lives in the speech itself, not behind it — a codec, a bad connection, an effect that was applied on purpose — and no noise remover, this one included, repairs it.
Standard is the right starting point for almost everything, and for a normal indoor recording it is the only one you will need. Gentle exists for the recordings where Standard goes too far, and the way to find out is to run it and listen. Nothing is lost by trying both: your file is never sent anywhere, and running it again costs a few seconds.
It does not even out the volume, does not make the file smaller, does not trim or split anything, and does not let you pick a different format to save in. It takes noise out and leaves everything else where it was. If you also need the file to be smaller, the audio compressor below does that job.
It is also not a repair tool. Clipping, dropouts, echo from a bare room and a microphone that was too far away are all damage in the voice itself, not noise sitting behind it, and none of them are improved here.
And it is built for speech and nothing else. Anything in a recording that is not a voice is treated as noise, so music handed to it does not come back cleaner — it comes back gutted. Use it on talking.
No. The work happens on your own device. You can switch your network off after it starts and it will still finish.
No, and nothing that claims to should be believed. It is very good at steady noise that sits behind the voice — a fan, a hiss, a mains hum, traffic. It is poor at other people talking, because that noise is speech too and there is no reliable way to tell whose voice is whose. Listen to the before and after in the box above before you download; that is why it is there.
Sustained applause in the pauses between sentences, mostly yes — we measured 30 to 40 dB of reduction once the clapping had run for about half a second. Single claps, mostly no: the first clap in our test passed through only 1.7 dB quieter, and a clap that lands while someone is talking is kept deliberately, because cutting it would cut the words around it. Run your file and use the before-and-after switch; that is the real answer for your recording.
No. A two-tone effect we measured came through 2.8 dB quieter — effectively unchanged — because an effect sounds nothing like the steady hiss the model strips. Music behind a voice fares no better, and music on its own comes back damaged, since everything that is not speech is treated as noise. Nothing on this site separates effects or music from a voice.
No. We ran a deliberately robotised voice through the filter and the output measured within a tenth of a decibel of the input: the tool decided the robotic voice was the voice, and protected it. That timbre is damage inside the speech — a codec, a bad connection, an effect — and a noise remover only takes things out from behind the speech. Nothing here repairs it.
Only when what makes it sound low-quality is steady noise. Hiss, hum and room rumble come out, and a recording whose problem was those genuinely sounds clearer afterwards. A muffled microphone, heavy compression artefacts or a voice recorded too quietly are not noise sitting behind the speech, and they come back exactly as they were.
Standard takes out as much noise as the tool can. Gentle blends the cleaned version with your original, so less noise comes out but less of the voice is touched. Gentle is worth trying on old, muffled or telephone-quality recordings, where a voice can be quiet and narrow enough that the tool mistakes parts of it for noise and thins it. If Standard sounds like it has chewed the speech, try Gentle.
No. The length and the number of channels are exactly what they were, and the file comes back at about the size it went in at. Making a recording smaller is a separate tool.
No. It only takes noise out. The speech is left at the level you recorded it, so a quiet recording comes back quiet — just with less behind it.
The same kind of file you dropped in, wherever that is possible: an MP3 comes back an MP3, a WAV comes back a WAV, a FLAC comes back a FLAC. A few formats cannot be written here — Opus, Vorbis, WMA and AMR — and those come back as MP3.
There is no fixed cap, because there is no upload. The practical limit is your device’s memory. A ten-minute recording takes a few seconds.
No watermark, no account, no trial, and no upgrade prompt on the download. The site is free and paid for by ads.