You can download VCTK Corpus(en multi speaker)/LJ-Speech(en single speaker) very easily via my repository.
set parameters
Almost parameters in params.py and teacher_params.py are same as params.py in my other repositories like VQ-VAE. If you modified params.py in AutoregressiveWavenet, you have to replace teacher_params.py with it to train student.
You can resume snapshot and restart training like below.(Now support AutoregressiveWaveNet only)
python train.py -r snapshot_iter_100000
Other arguments -f and -p are parameters for multiprocess in preprocessing. -f means the number of prefetch and -p means the number of processes. I highly recommend to modify -f to large number like 64. If GPU-Util is stil low, modify -p to large number like 8.