How to train WaveRNN

[dr-jaieun]

Your WaveRNN repository(master-barnch) excludes preprocess.py modules, how can I train the WaveRNN model? @begeekmyfriend

[begeekmyfriend]

It is quite easy to understand that you only need two directories called gta (or mel for ground truth) and quant (for wav clip) to each speaker for DataLoader according to dataset.py. The file format is all npy and the sizes of the first (or maybe last that I do not remember) dimension of npy must match with each other. For instance, you might just copy the wav clips into quant directory generated in preprocessor.py in tacotron2. And the GTA mel can be generated through scripts/gta_synth.sh.

[dr-jaieun]

@begeekmyfriend It can be considered that we can use the results(wav, mels, gta) generated through the preprocessing of this tacotron2 repository. It's a simple issue, but it can be helpful to automate the pipeline between the two repository directory structures. I will try and PR. Thank you for the quick answer.