Uses spectral subtraction — the same technique behind Audacity's Noise Reduction effect — to learn a noise profile from a quiet section and subtract it from the rest of the file.
The Noise Reducer removes steady background noise — hiss, hum, fans, air-conditioning, room tone and tape noise — from recordings, using spectral noise reduction that runs entirely in your browser.
It learns what the noise looks like (from your selected sample or automatically from the quietest parts of the file), then lowers those frequencies wherever they are not masked by wanted sound. Smoothing over time and frequency avoids the “musical noise” — warbling, bird-like artefacts — that makes cheap noise removal sound worse than the noise.
The file is split into 2048-sample frames with 75% overlap and a Hann window, and transformed with an FFT. A noise profile (average power per frequency bin) is built from your selection, or from the quietest 12% of frames when in auto mode. For each bin a Wiener-style gain √(1−1/SNR) is computed using a profile scaled by an over-subtraction factor of 1–4.5 (your strength setting), then smoothed across three adjacent bins and over time with fast attack and slow release, and floored at −8…−30 dB. Frames are transformed back with weighted overlap-add.
The report uses a robust noise-floor estimate — the 10th percentile of 50 ms RMS levels — before and after, so you see the real improvement.
Remove fan and room noise from a home recording.
Reduce tape hiss and hum on digitised cassettes.
Clean up wind-free ambient hiss on interviews.
Lower amp hiss before mixing.
| Algorithm | STFT spectral gating with Wiener gain, 2048 FFT, 75% overlap |
|---|---|
| Strength | 0–100% (over-subtraction 1–4.5, floor −8…−30 dB) |
| Profile | Selected sample or automatic (quietest 12%) |
| Measured result | Noise floor before/after in dB |
Yes, best on steady noise. Random bursts (traffic, keyboard clicks) are not steady and will not be removed.
No. It targets noise, not reverb.
Never — it is processed on your device.
Sudden on/off gain decisions per frequency bin; smoothing in time and frequency prevents it.
It measures the average frequency 'fingerprint' of the noise sample you select, then subtracts that fingerprint from every part of the file — the classic spectral-subtraction technique used by professional tools like Audacity.
It significantly reduces steady, consistent background noise (hiss, hum, fan noise, room tone). It won't remove noises that change over time (a dog barking, traffic) — those need per-instance editing instead.
That's a known side effect of spectral subtraction at higher strengths, sometimes called 'musical noise.' Try a lower strength setting, or pick a noise sample that's more representative of the noise throughout the file.
This tool runs entirely in your browser using the Web Audio API. Nothing is uploaded, stored or shared — there is no server to send your file to, so it works offline once the page has loaded.
Be the first to share your experience with this tool — your feedback helps others.