Official implementation of SpeechSplit 2.0: Unsupervised speech disentanglement for voice conversion Without tuning autoencoder Bottlenecks.

The audio demo can be found here.

The WaveNet vocoder is the same as in AutoVC. Please refer to the original repo to download the vocoder.

To run the demo, first create a new directory called models and download the pretrained models and the WaveNet vocoder into this directory. Then, run demo.ipynb. The converted results will be saved under result.

Download the VCTK Corpus and place it under data/train. The data directory should look like:

  data
    |__train
    |    |__P225
    |        |__wavfile1
    |        |__wavfile2
    |        ...
    |    |__P226
    |     ...
    |__test
         |__p225_001.wav # source audio for demo
         |__p258_001.wav # target audio for demo

NOTE: The released models were trained only on a subset of speakers in the VCTK corpus. The full list of speakers for training is encoded as a dictionary and saved in spk_meta.pkl. If you want to train with more speakers or use another dataset, please prepare the metadata in the following key-value format:

speaker: (id, gender)

where speaker should be a string, id should be a unique integer for each speaker(will be used to generate one-hot speaker vector), and gender should either be "M"(for male) and "F"(for female).

To generate features, run

python main.py --stage 0

By default, all generated features are saved in the feat directory.

To train a model from scratch, run

python main.py --stage 1 --config_name spsp2-large --model_type G

To finetune a pretrained model(make sure all pretrained models are downloaded into models), run

python main.py --stage 1 --config_name spsp2-large --model_type G --resume_iters 800000

If you want to train the variant with smaller bottleneck, replace spsp2-large with spsp2-small. If you want to train the pitch converter, replace G with F.