Audio Noise Reducer
Remove hiss, hum and room noise from a recording with spectral gating — on device.
About this tool
Drop a recording; the quietest moments become a noise print; every slice of the spectrum sitting at that level is turned down while louder sound passes. Adjust strength and the maximum reduction, listen, download WAV. Nothing is uploaded.
How spectral gating works, and the over-processing trap
Steady background noise — air conditioning, laptop fans, electrical hum, tape hiss, the rumble of a room — has a signature: the same spectrum, all the time. Speech and music don't. The tool splits the audio into short overlapping windows, measures each window's spectrum with a Fourier transform, learns the noise signature from the quietest tenth of the file, and applies a per-frequency gate: bins that stay near the noise level are attenuated (down to your chosen floor, −12 to −30 dB), bins that rise above it are kept. This is the technique in Audacity's Noise Reduction and in most podcast tools, running entirely in your browser. Honest limits: it removes noise that is constant; a barking dog, traffic, keyboard clicks or a second voice are not “noise” to the algorithm and stay. And there is a cliff: push strength and reduction too far and speech takes on the watery, robotic “musical noise” artefact every over-processed podcast has. Start at the defaults and listen. Normalise the result to bring the level up, trim the pauses, then encode to MP3.
Frequently asked questions
What kinds of noise does it remove?
Steady noise: hiss, hum, fans, air conditioning, room rumble, tape noise. Not intermittent sounds like dogs, traffic, clicks or other voices — those need editing, not gating.
Why does the voice sound watery?
Too much reduction — the classic artefact of spectral gating. Lower the strength or choose the gentle −12 dB floor. A little remaining noise sounds better than a robotic voice.
Does it need a noise sample?
No — the noise print is learned automatically from the quietest tenth of the recording. A second or two of 'silence' somewhere in the file helps it.
Is the audio uploaded?
No — the analysis runs in your browser; a few seconds for a three-minute file.