This repo contains the training code for deep neural pitch extractor for Voice Conversion (VC) and TTS used in StarGANv2-VC and StyleTTS. This is the F0 network in StarGANv2-VC and pitch extractor in StyleTTS.

Please specify the training and validation data in config.yml file. The data list format needs to be filename.wav|anything, see train_list.txt as an example (a subset of VCTK). Note that you can put anything after the filename because the training labels are generated ad-hoc.

Checkpoints and Tensorboard logs will be saved at log_dir. To speed up training, you may want to make batch_size as large as your GPU RAM can take.

Since both harvest and dio are relatively slow, we do have to save the computed F0 ground truth for later use. In meldataset.py, it will write the computed F0 curve _f0.npy for each .wav file. This requires write permission in your data folder.

In meldataset.py, the F0 curves are computated using PyWorld, one with harvest and another with dio. Both methods are acoustic-based and are unstable under certain conditions. harvest is faster but fails more than dio, so we first try harvest. When harvest fails (determined by number of frames with non-zero values), it will compute the ground truth F0 labels with dio. If dio fails, the computed F0 will have NaN and will be replaced with 0. This is supposed to occur only occasionally and should not affect training because these samples are treated as noises by the neural network and deep learning models are kwown to even benefit from slightly noisy datasets. However, if a lot of your samples have this problem (say > 5%), please remove them from the training set so that the model does not learn from the failed samples.

Data augmentation is not included in this code. For better voice conversion results, please add your own data augmentation in meldataset.py with audiomentations.