Take steady background noise out of a recording by first measuring what the noise alone sounds like. A fan, an air conditioner, a preamp’s hiss: anything that stays the same from start to finish can be learned once and subtracted everywhere.
Learn the noise, then remove it
Noise reduction works in two steps. The first measures a noise profile, the level of the noise in each of 96 frequency bands, from a stretch of the file that holds nothing but noise. The second runs through the whole recording and turns down every part of the spectrum that is no louder than that profile.
When a file loads, the quietest stretch between 0.5 and 2 seconds long is found and learned from automatically. Stretches of pure digital silence are skipped, because zeros say nothing about the room. The panel above the sliders says where the current profile came from.
Choosing the noise yourself
The automatic choice is the quietest part of the file, which is usually room tone. A recording that never pauses can push it onto a soft syllable instead. In that case, play the file, find a gap with only the background in it, drag across it on the waveform, and press Use selection as noise.
Half a second of clean noise is enough. Longer is fine, as long as nothing else sounds in it. A breath, a chair creak, or the tail of a word inside the selection becomes part of the profile and is taken out of the rest of the file too.
Reduction, sensitivity, and smoothing
Reduction is how far the noise is pulled down, from 0 to 40 dB. The default of 12 dB leaves a little room tone in the pauses, which sounds more natural on voice than dead silence. Higher values push the gaps between words toward silence.
Sensitivity decides what counts as noise. At 50% a sound has to rise a little above the profile’s level to be kept. Raise it and more of the quiet detail goes with the noise; lower it and some of the noise survives.
Smoothing averages the reduction across neighbouring frequencies. Spectral subtraction with no smoothing leaves isolated blips of noise that come and go, a chirping artifact called musical noise. Raising Smoothing blends those blips into the reduced background. Frequencies that sit well above the noise keep their own level either way.
Profile vs adaptive denoising
A learned profile is exact for noise that does not change. A hum that drifts, traffic going past, or a room whose noise changes halfway through a take is better served by the Audio Denoiser, which keeps re-estimating the noise as the file plays.
Example: a voice memo with laptop fan noise
The first second of the recording is the fan alone, before the speaker starts. That is the quietest stretch, so it is learned automatically. With Reduction at 15 dB, Sensitivity 50%, and Smoothing 50%, the fan drops well below the voice in the pauses, and the voice keeps its treble because it sits far above the fan’s level in those bands.