Audio Noise Reducer
Reduce approximately stationary background noise using a bounded browser-local spectral noise profile and overlap-add processing.
How it works
Build 115 estimates a frequency-by-frequency noise profile from the beginning of the file, then applies a conservative spectral gate using 1024-point Hann-window FFT frames with 50% overlap-add reconstruction.
Prepare the noise profile
The beginning of the file should contain mostly the background noise you want to reduce. If speech, music or transients dominate the profile section, the reducer may suppress similar frequencies later.
Limitations
This is a deterministic stationary-noise reducer, not AI source separation, dialogue isolation, dereverberation, forensic restoration or a studio spectral editor. Aggressive settings can create musical-noise or watery artifacts. It does not learn from remote services.
Recommended workflow
Start around strength 1.5–2.5 and a floor of −18 to −30 dB. Compare with the original and prefer a smaller reduction when artifacts become audible.
Privacy & output
Processing remains browser-local. Export is 16-bit PCM WAV; metadata and artwork are not preserved.
FAQ
Can it remove voices? No. Can it clean traffic that changes constantly? Only partially; non-stationary noise is outside this simple profile model.