Remove Background Noise, Free nothing stored

How the noise model tells a voice from the noise around it

The cleaner is an open-source, BSD-licensed noise-suppression model, compiled to WebAssembly and loaded from this site as a file of about 1.4 MB. It listens in 10 ms frames at 48 kHz, so the sound is resampled to that rate first. Each channel gets its own model state; stereo stays stereo, and surround is folded down to two channels.

In every frame the model keeps what sounds like speech and turns the rest down. That suits steady and broadband noise: hiss, hum, fans, air conditioning, traffic and room noise. Each frame also gets a voice score, which feeds the Voice found figure on the result.

What the noise removal percentage changes

Noise removal is a mix, not a preset. At 90% the output is that share of cleaned sound plus the rest of the original recording, lined up sample for sample, so the two never smear or double when combined.

Full strength removes the most, but on music or in a very noisy room it can sound thin or underwater. Turning it down lets a little of the room back in, which often makes a voice sound natural again. The Noise removed field on the result repeats the strength you used, so two runs are easy to compare.

Cleaning a video’s sound while the picture stays as recorded

With a video, only the sound is decoded, cleaned and encoded again. The picture is copied into the new file without re-encoding, and the result’s Picture field reads Original, not re-encoded. MP4 and MOV videos come back as MP4 with AAC sound, or Opus when the browser has no AAC encoder; WebM stays WebM with Opus sound.

Timing is corrected twice. The model’s own 20 ms lag is removed, and so is the silent lead-in some encoders add, which is measured in your browser before the run. The cleaned track ends where the original did. Only the main video and audio tracks are kept, and a picture format this browser cannot copy stops the run with a message.

Saving cleaned audio as M4A or WAV

For an audio file, Format offers M4A or WAV, starting at M4A. M4A holds compressed AAC sound at 96 kbps per channel. WAV holds 16-bit samples with no compression, which suits further editing. Both are written at 48 kHz, the rate the model works at. MP3 is accepted as input but never written.

If the browser cannot save M4A, the run stops and suggests WAV. Older browsers that cannot decode audio on the page read the whole file at once instead; that route saves WAV only and takes files up to 300 MB.

Sounds the voice cleaner leaves in place

Because the model keeps whatever sounds like speech, other voices tend to stay too: a TV, a crowd or a second person in the room. Echo from a bare room is made of the voice itself, and this tool has no step that removes it. A closer microphone or a softer room fixes echo at the source.

Music is not speech, so under a voice it is thinned out along with the noise, and a song on its own can lose much of its sound. If the file barely sounds like a voice, the result says that little speech was found. That note is a hint that the file is not a voice recording.

Your recording is cleaned on this device

The file is read from your device, cleaned in the browser tab and assembled in memory as a new file. It is not uploaded for noise removal. The page loads its own code and the noise model, and a failed run can send an error report with the file’s size and type, never the sound.

The working budget is 15 minutes or 1 GiB per run on a phone and 60 minutes or 4 GiB on a computer. These are upper limits rather than promises: an older phone can run out of memory sooner. Trim a long recording into parts first if it does not fit.

Where a cleaned recording goes next

The free Audio to Text tool can turn the cleaned take into SRT subtitles or a transcript. When a line was fluffed or never recorded, the paid voice studio can learn your voice from a clean sample and speak new text in it.

For a wider audience, the paid dubbing studio can voice the cleaned video in other languages. Noise in the picture, such as grain, is a separate job for the paid video enhancer.

How do you use the noise remover?

1. Add the noisy recording

   Choose a MP3, M4A, WAV, MP4, MOV or WebM file: a voice memo, a podcast take, an interview or a talking-head video.

2. Set strength and format

   Leave Noise removal at 90% for a noisy room, or lower it for a lighter touch. Format applies to audio files; a video keeps its own container.

3. Run it and listen

   Press Remove noise, then play the result next to the original. Check Voice found, Format, Size and, for a video, Picture.

4. Save your result

   2 free downloads, 8 more after sign-in.

When is the free noise remover enough?

Noise Remover: questions and answers