Skip to content
Cutshed
Menu

Audio · 06

Noise remover

RNNoise is a small speech denoiser. It turns down steady hiss, fan noise, and room tone on a voice. It is a poor fit for music.

RNNoise is a small speech denoiser (about 0.1 MB). It was trained on voice, not on music beds. A noisy interview gets better. A song can lose cymbals and room. The result is WAV.

How to use it

The page decodes your file, resamples each channel to 48 kHz, and runs RNNoise in 480-sample frames. The cleaned audio is resampled back and saved as WAV. The wasm module is about 110 KB and loads on the first run.

RNNoise was trained on speech. A song, a guitar, or a room full of music will sound watery or dull. Use it on interviews, voiceovers, and calls. Leave music alone, or split the vocal first and denoise only that stem if you must.

It does not remove a door slam, a cough, or a siren that happens once. Those are not stationary noise. Edit them out.

Tips

  • Give it a take where the noise is steady. A laptop fan is the case it was built for.
  • If the voice turns robotic, the noise was already louder than the speech. Record closer next time.
  • Denoise before you boost, or boost only a little first. Huge gain lifts the noise floor into the model’s working range and the voice gets chewed.

FAQ

Which model is this?

The built-in RNNoise model from the xiph.org project, compiled to WebAssembly by the Jitsi rnnoise-wasm package. No weights are fetched from a server beyond that small wasm file.

Will it clean a song?

Expect damage. Music is not the training set. Use it on speech.

Does the audio leave the browser?

No. The wasm file is the only download, and it is the denoiser, not your recording.

License

RNNoise is BSD-3-Clause (Xiph.Org, Jean-Marc Valin). The @jitsi/rnnoise-wasm package that wraps it for the browser is Apache-2.0. Both allow commercial use. The license covers the denoiser, not the recording you run through it.