This is an unofficial implementation of the paper One-shot Voice Conversion by Separating Speaker and Content Representations with Instance Normalization modified from the official one.

• Python >= 3.6
• torch >= 1.7.0
• torchaudio >= 0.7.0
• numpy >= 1.16.0
• librosa >= 0.6.3

The main difference from the official implementation is the use of a neural vocoder, which greatly improves the audio quality. I adopted universal vocoder, whose code was from yistLin/universal-vocoder and checkpoint will be available soon. Besides, this implementation supports torch.jit, so the full model can be loaded with simply one line:

model = torch.jit.load(model_path)

Pre-trained models are available here.

The code preprocess.py extracts features from raw audios.

python preprocess.py <data_dir> <save_dir> [--segment seg_len]

• data_dir: The directory of speakers.
• save_dir: The directory to save the processed files.
• seg_len: The length of segments for training.

python train.py <config_file> <data_dir> <save_dir> [--n_steps steps] [--save_steps save] [--log_steps log] [--n_spks spks] [--n_uttrs uttrs]

• config_file: The config file for AdaIN-VC.
• data_dir: The directory of processed files given by preprocess.py.
• save_dir: The directory to save the model.
• steps: The number of steps for training.
• save: To save the model every save steps.
• log: To record training information every log steps.
• spks: The number of speakers in the batch.
• uttrs: The number of utterances for each speaker in the batch.

You can use inference.py to perform one-shot voice conversion. The pre-trained model will be available soon.

python inference.py <model_path> <vocoder_path> <source> <target> <output>

• model_path: The path of the model file.
• vocoder_path: The path of the vocoder file.
• source: The utterance providing linguistic content.
• target: The utterance providing target speaker timbre.
• output: The converted utterance.

Please cite the paper if you find AdaIN-VC useful.

@article{chou2019one,
  title={One-shot voice conversion by separating speaker and content representations with instance normalization},
  author={Chou, Ju-chieh and Yeh, Cheng-chieh and Lee, Hung-yi},
  journal={arXiv preprint arXiv:1904.05742},
  year={2019}
}