[nitrogen]

I wonder how a hybrid model would sound, where the net generates parameters for a parametric synthesis algorithm (or a common speech codec) instead of samples, to reduce CPU costs.