Remove background noise

Traffic through a window. A fan you stopped hearing an hour ago. The hum a room adds to everything recorded in it. This page is about what can be taken back out of a recording, what cannot, and how to find out which one you have — in about a minute, for free, without sending us the file.

What the engine is actually doing

The noise removal is DeepFilterNet 3, a speech enhancement model. It has no rule for "traffic" and no setting for "air conditioning". What it has is a learned idea of what human speech looks like moment to moment, and it suppresses what does not fit that idea.

That is the whole thing, and it is worth understanding, because it changes the question. The question is not which noise you have. It is how separable your voice is from it.

It runs in short frames as the audio goes past rather than studying the whole file, which is why a 30 second preview takes seconds rather than minutes — on an iPhone included. And because the engine is WebAssembly running in your own browser, those seconds are spent on your device.

Where it works, and where it will struggle

What follows is reasoning from how the model works, not a test we ran on a file like yours. We have taken exactly one recording end to end and published it — the street interview on the home page, traffic and all. Everything else would be us guessing on your behalf.

• Easier: a continuous, unstructured sound sitting behind a clear voice. There is a lot of signal to tell the two apart.
• Harder: other people talking. This is a speech model, so it keeps speech — including the speech you did not want. Background conversation is close to the worst case, not the easy one.
• Harder: anything sitting on top of the voice in both time and frequency at once, because once the two overlap that closely there is not much left underneath to reconstruct.
• Out of scope: damage. Clipping, a microphone that was knocked, a voice already buried under the noise floor. Noise removal takes things away; it cannot put back what was never recorded.
• Unknown: room echo. Echo is your own voice arriving late, not something added on top of it, and we have not measured how much of it this model removes. Assume it stays, and check on the player before you plan around it.

We are not going to publish a table of noises we beat. The honest answer for your recording is the A/B player on your recording, which is why the preview is free and why it takes under a minute.

Three presets, and no wrong answer

Once the preview is ready you can switch between three settings as often as you like. Switching re-runs the noise removal but not the rest of the work, so it comes back in seconds.

• Balanced — the model's own defaults. Recommended for most files.
• Strong — pushes harder, and may sound watery or robotic.
• Clear Voice — less aggressive, keeps more of the room.

Noise removal is lossy: it is always trading some of the voice for some of the noise. If Balanced already sounds right, stop there. If it does not, the fastest way to find out whether Strong is an improvement or a disaster is to press it and listen.

Nothing is uploaded

There is no upload step to opt out of. Your browser opens the file locally, and the engine — about 9 MB of WebAssembly — is fetched to you, once, after you pick a file. Landing on this page does not download it. Your audio is never used to train anything either, and not as a promise: we could not, because we never receive it. The details are in our privacy note.