TeeJayBaker / PolyDDSP

Polyphonic generalisation of DDSP
16 stars 0 forks source link

Design source separation architecture #1

Open TeeJayBaker opened 9 months ago

TeeJayBaker commented 9 months ago

Copy the basic pitch convolution architecture taking in stacked harmonic CQT and freq encoding to output both a individual CQT and an MFCC

Target multiple objective functions, one being MSSL on the output spectrogram, and 2 being MFCC loss against the correct individual voice timbre.