Pull down steady background noise, room tone, and hiss without recording a separate noise profile first.
Adaptive floor, not a captured profile
Some denoisers ask for a stretch of pure noise before they can work. This one tracks a floor per frequency band as the file plays: the estimate drops instantly whenever a band goes quieter than what it has already seen, and climbs back only slowly otherwise. That asymmetry is what keeps a loud passage from dragging the floor up and defeating the gate later in the file.
Because the floor is a leaky minimum, it needs at least one quiet moment, a pause, a breath, room tone before the take starts, to know where silence sits. A file that never drops in level gives it nothing to measure against.
Bands, not bins
The reduction is computed across 28 log-spaced frequency bands rather than every individual FFT bin. Gating whole bands at once, and smoothing each band’s gain across roughly 120 ms, is what keeps the result from chattering: reducing bin by bin makes the noise floor itself audible as a fluttering artifact usually called musical noise.
Reduction, sensitivity, and preserve highs
Reduction sets how far a band is pulled down once it is judged to be noise, up to 40 dB. Sensitivity sets how close to the floor a band has to sit before that judgment applies. Preserve highs tapers the reduction above 3 kHz, since gating there is the most audible place for chatter to show up, on cymbals, sibilance, and room air.
Example: a voice recording with fan noise
Reduction 20 dB, sensitivity 55%, preserve highs 40%.
The fan’s steady hum sits in a narrow, unchanging band and gets tracked and gated within the first second or two. The voice, which moves across a much wider range and never sits still, stays mostly above the floor and passes through with the treble intact.