Tools › Audio

Clean up a voice recording

Remove background noise from audio: take hiss, hum and room noise out of a podcast, lecture or voice memo, shorten long pauses and cut the “um”s. No upload.

How it works

  1. Noise reduction is spectral gating, the method behind Audacity’s noise removal: the recording is split into 2048-sample frames, the noise floor of every frequency band is estimated from the quietest parts of the recording itself, and each band is turned down where it does not rise above that floor. Hiss, hum, fans and air conditioning go down; the voice stays.
  2. Pauses are found on the loudness envelope: every stretch quieter than the speech level by 30 dB and longer than your limit is shortened to the length you choose, with cross-fades so nothing clicks.
  3. Fillers are found by Whisper running on your device with word timestamps. Only sounds without meaning are cut: “uh”, “um”, “hmm” and their equivalents in 16 languages. “Like” and “you know” are left alone, because cutting them changes sentences.
  4. The result is saved as MP3, M4A, Opus, OGG, FLAC or WAV at the original sample rate.

Questions

Is the recording uploaded?

No. Decoding, noise reduction, cutting and encoding all run in this tab. If you turn on filler removal, the speech model is downloaded once and then works offline.

Will it remove a barking dog or traffic?

Not well. Spectral gating removes steady noise: hiss, hum, fans, room tone. A sudden loud sound is not noise to it. For those, cut the moment out with the Audio Cutter.

Does it work on music?

It is made for speech. On music, high settings make reverb tails and cymbals sound watery; keep the strength low.

Can I clean the audio of a video?

Yes, drop the video and you get the cleaned soundtrack as an audio file. Putting it back into the video is on the way.

Which languages are the fillers found in?

English, Spanish, Portuguese, Indonesian, Russian, Japanese, Arabic, Hindi, Vietnamese, Turkish, German, French, Italian, Korean, Polish and Ukrainian. The language is detected from the recording.