Clean up a recording
This free background noise remover learns the hum, hiss or fan from a quiet passage of your own recording and subtracts it across the whole file — then plays you the before and after so you can judge it yourself. It runs entirely inside this tab, and the recording is never uploaded: the tools on this site never upload your audio.
Ready.
How it works
The noise profile comes from the quietest passage in your file — the gap before you start speaking, or between sentences. There is no trained model here and nothing about your voice is sent anywhere to be analysed. If the recording has no quiet stretch at all, the page says so and refuses, because there is nothing to learn the noise from.
Every frequency band keeps a floor of gain, so no band is ever silenced. That is the defence against the warbling — musical noise — that cheap denoisers leave behind, and it costs reach on purpose: the subtraction stops at 12 dB however hard you push the slider. A noisier recording does not simply come out clean, and this page will tell you how far it actually got.
Subtraction reaches noise that sits underneath speech, which gating cannot. What it cannot do is invent the part of your voice the noise was covering. When speech is very close to the noise, the honest answer is to record again in a quieter place, and no setting here changes that.
The floor is measured before, and again from the encoded MP3 you download — not from the samples handed to the encoder. The encoder is part of what you get, so it is part of what gets checked.
Recording for publication?
Audiobook and podcast platforms publish a limit for background noise, and cleaning a file up afterwards is the second-best way to meet it. These tools take it further, and none of them uploads anything either.
The room test measures your space against the level publishers require, and the microphone test tells you whether the input itself is set up right.
The chapter checker measures a finished file against ACX, Spotify, Google Play and Apple and corrects what is correctable — this same subtraction is one of the stages it uses, aimed at a stated limit rather than at a slider.
Straight answers
Choose the file and this page finds the quietest passage in it, measures the noise there, and subtracts that spectrum across the whole recording. You get the noise floor before and after, the cleaned MP3, and the first twenty seconds of both takes to compare by ear. Nothing is uploaded; the file is read inside the browser tab.
Partly, and that is the difference between this and a noise gate. A gate can only quieten the gaps between words; subtracting in the frequency domain reaches noise sitting under speech too. What it cannot do is restore the part of your voice the noise was covering, so a recording where speech is very close to the noise comes out cleaner but not clean.
Every frequency band keeps a floor of gain here, so no band is ever silenced. That floor is what prevents the warbling artefact — musical noise — that heavier denoisers leave, and it caps how far the subtraction reaches. The page reports how far the floor actually moved rather than the number you asked for.
Then the page refuses instead of guessing. Spectral subtraction needs a passage with no speech in it, even half a second, to learn what the noise is. With nothing to learn from there is no profile to subtract, and a tool that carried on would be removing something it never measured.
The method is the same family — learn a noise profile from a selection, subtract it spectrally. The differences are that this runs in a browser tab with nothing to install and nothing uploaded, that the profile is found automatically from your file's own quietest passage, and that the result is re-measured from the encoded MP3 you download rather than from the samples before encoding.