-
Fix clipping
- If wet audio has values over 1, normalize wet and dry accordingly --> done
-
Understand and apply spectral loss --> done
- plot spectrums of x and y --> to check
- plot distance spectrums --> to check
-
Fix batches name in AudioDataset use couples or similar --> done
-
Think about mono and stereo aglomeration in preprocessing
- Do not throw away the mono samples --> done
-
Understand and apply pqmf
- do decomposition --> done
- listen to audio after decomposition --> DOne
- Compare the shapes --> Done
- do reconstruction --> Done
-
Do the block results list directly in the encoder achitecture --> Done
-
Normalize clipping by pair not individually ! --> Done
-
Do spectral loss plots locally rather than on tensorboard. --> Done
-
Also plot signals of the PQMF. time domain/ spectral/ --> Done
-
plot both in time and then in fft. --> Done
-
Have a properly wet signal --> Done
-
Have a right plot with right colors, make code more minimal --> Done
-
Investigate log distance without adding small value
-
Clean spectral notebook
-
Also do PQMF on white noise --> Done
-
Understand Micha notebook about PQMF
-
write in big our issue
-
strip down notebook but keep formulas and important shut
-
Smoothen the spectrums with window averaging