The model would be good, if not for these wheezes and chewing in covers. But it's probably the algorithm's fault, because even in other good vocal models it's the same. I wonder, if I don't feed the model full-fledged a cappellas during training, but cut them into phrases and subtypes and load them like this, will the wheezing and chewing stop? Or will I still have to load phrases at different pitches separately (having done some wizardry in the same melodyine)?