Tacotron 2 with ParallelWaveGAN. Next step

Yes, didn‘t try myself yet, but it is possible by using option --restore_path for training,

Indeed it was possible. Achieved same or better results from pretrained model.

If the dataset is clean enough, training PWGAN with the real spectrograms would be sufficient.

To be honest, my dataset is anything but clean. I am thinking about replacing bad audio with better as I go along. Is this a good idea or should i retrain from start?

Clean data is always the right choice.

haha of course, but it depends on what possibilities you have of achieving that data :slight_smile:

can you share a sample for us to see how good your data is?

Sure

This is in swedish, but other sets are similar quality

is it something private you train or might be publicized ?

Just a private thing :slight_smile:

would you say audio quality is good enough?