ihateaudio

Background Noise Remover

Take the room out and leave the voice.

Drop an audio file here

or paste one from your clipboard

MP3 · WAV · M4A · AAC · OGG · FLAC and moreYour file stays on your device. Always.

How to use it

  1. Drop the recording in. There is no model to download for this one: it is small enough to ship with the site, so it starts working straight away.
  2. Press Clean. The whole file is processed about eighty times faster than real time, so even a long interview is done in seconds rather than minutes.
  3. Listen, and compare. The cleaned version appears as its own player under the controls, with the original still in the player above, so you can play one against the other and hear exactly what left. Full strength is not always best: on a recording where the room is part of the atmosphere, leaving a little of the original in sounds more natural.
  4. Download when it sounds right. If this is going on to a transcriber, run it through cleaned first: speech recognition is noticeably more accurate once the hiss is gone.

How a 150 KB model does this

Most noise reduction works by subtraction: measure the noise during a silent passage, then subtract that profile from the whole recording. It works, and its failure mode is famous: the musical-noise artefact, a shimmering of little tones where the subtraction left uneven residue behind.

RNNoise does something different. It splits the signal into a couple of dozen frequency bands and, for every 10-millisecond frame, asks a small recurrent model how much of each band is voice. The answer is a gain per band, and the gains are applied smoothly. Because the model carries state from one frame to the next, it knows whether it is in the middle of a word or in a gap, which is the information a static noise profile does not have.

It is small because it only ever had to output those couple of dozen gains rather than reconstruct a waveform, and because it was built to run on modest hardware in real time. That constraint is why it can ship inside a web page instead of behind a download.

Where it helps most

Interviews and voice notes recorded in ordinary rooms are the best case. So is anything you intend to transcribe: speech recognition degrades quickly with background noise, and cleaning first measurably improves the transcript.

It also pairs well with the silence remover. Cleaning first and then removing silence gives a better result than the other way round, because the silence detector is looking for passages below a level threshold, and it is far easier to pick those out once the noise floor has come down.

When to leave it alone

Anything where the background is content. Ambience recorded deliberately, a live room you wanted the sound of, a field recording: this will strip exactly the thing you went there for.

Music, as above: it was trained on speech and will treat sustained notes with suspicion. And anything you may need as evidence, where the recording should stay as it was captured rather than as it sounded best.

Questions

What kind of noise does this remove?
Steady background noise: air conditioning, computer fans, traffic hum through a window, tape hiss, the general roar of a room. It is very good at anything continuous and unchanging. It is much less good at sudden noises (a door slamming, a cough, a chair scraping) because those look like part of the speech to a model trained to separate voice from steady background.
Why is there nothing to download here when the other AI tools need 40 or 64 MB?
Because this model is tiny. RNNoise is about 150 KB, which is smaller than most photographs, and it ships with the site like any other asset. It manages that by being narrow: it does one job, on speech, at one sample rate, and it was designed in an era when running an AI model on a phone meant making it small enough to fit. The separation models are large because they are doing something far harder.
It made the voice sound thin or underwater.
Turn the amount down. The model decides, band by band and moment by moment, how much of what it hears is voice, and where it is uncertain it attenuates. On a recording where the noise and the voice overlap heavily it removes some of the voice with the noise, and the result is that hollow, processed quality. Mixing back 20 or 30 per cent of the original usually restores the body without bringing the noise back.
Can it remove music from behind speech?
Not well, and it is a different problem. This is trained to tell voice from steady noise, and music is neither steady nor noise. If there is music under the dialogue, the vocal remover does a far better job, because separating a voice from an arrangement is exactly what it was built for.
Does it work on singing?
Less well than on speech. RNNoise was trained on speech, so it has learned what a talking voice looks like, and sustained sung notes sometimes get treated as tonal noise and attenuated. For music, use the separation tools instead.
Why does the file come back at a different quality?
It should not, but there is a small unavoidable round trip: the model runs at 48 kHz, so a 44.1 kHz recording is resampled up, processed, and resampled back. Both conversions use the browser’s own resampler with a proper filter, so the loss is far below anything audible, but it is not literally bit-identical in the passages where nothing was removed.