Open jackguo6 opened 11 months ago
It looks a bit like dithering noise.
It looks a bit like dithering noise.
We still don't know how it came about,any suggestions?
I've checked on my side, and it's not dithering, it's noise created by demucs during training, but maybe that seems to disappear when model is trained enough (as baseline models don't have that noise ?)
Here's a spectrogram with that obvious "synthetic" high freq noise with a model trained only with a few epochs:
btw, this repo is inactive as Demucs developpement was stopped and as A. Defossez has left Meta, the new demucs main repo is (for major bug reporting) : https://github.com/adefossez/demucs
❓ Questions
The model we trained ourselves has high-frequency noise in the audio data of the tracks, each track of data has similar noise. Adding additional data makes it more obvious.Any useful suggestions?