Closed lucianodato closed 7 years ago
If possible this should be frequency dependent.
Essentially the idea is to scale down the noise profile when a transient is present in the signal. it should be frequency dependent otherwise sounds bad. This should be activated by default some hints of a better method are in Adaptive Noise Reduction for Real-time Applications which propose a onset detection based on phase information. More info in Bello onset paper. Spectral flux is essentially crap since it's not frequency dependent and noise still appear. Other approach is to use a multiresolution STFT as proposed by Lukin in Adaptive Time-Frequency Resolution for Analysis and Processing of Audio but something like that would be a part of the STFT transform
Implement something to preserve transients. Some clues in Adaptive Noise Reduction for Real-time Applications paper