How to use it
- Open a recording (ideally with a second or two of silence in it).
- Leave strength at 1.5× and press "Remove noise".
- Listen. If noise remains, undo and try 2×. If the voice sounds watery, go lower.
- Download.
How spectral noise reduction works
The file is cut into short overlapping frames and each is converted to frequencies with an FFT. The quietest 10% of frames are averaged into a noise fingerprint. Every frame then has that fingerprint subtracted, frequency by frequency, with a floor so nothing is removed completely, and is converted back to sound.
It works well on constant noise and poorly on changing noise such as music, voices in the background, or traffic. It does not use AI. For really noisy recordings the best fix is still upstream: get closer to the mic and turn off the fan.
Questions
Is my audio uploaded anywhere?
No. The file is decoded, processed and encoded by your own browser using the Web Audio API. Nothing is sent to a server, which is also why it works on large files without a queue.
Why does my voice sound robotic after noise removal?
Strength is too high, so parts of your voice that resemble the noise are removed too, leaving "musical noise" artifacts. Undo and lower the strength.
Does this use AI?
No, it is classic spectral subtraction, the same principle as the noise reduction in Audacity. It runs entirely in your browser.