Kaizhi Qian

@liveroomand Yes in our case. but you can design your own speaker encoder or just use onehot embedding

@miaoYuanyuan For other dataset, you need to tune the parameters of the conversion model instead of the parameters of the feature.

@miaoYuanyuan If you change the parameters of features, you will need to retrain the wavenet-vocoder as well.

Please refer to the data preparation code for details

You don't need to train them at the same time.

You can simply replace G with P along with some other minor modifications.

All preprocessing steps are in the code, except trimming silence. But I don't think they will make any fundamental difference. Your loss value looks fine.

The train.pkl is intended for training.

For testing, please refer to this issue #108

the .pkl is not a format, it is just a suffix of the filename. You can name it whatever you like such as .abc, .qaz, or .wsx, etc. To save...