• This code is an implementation of AutoVC. The algorithm is based on the following paper:

  • Qian, K., Zhang, Y., Chang, S., Yang, X., & Hasegawa-Johnson, M. (2019). AutoVC:Zero-shot voice style transfer with only autoencoder loss. arXiv preprint arXiv:1905.05879.

  • Qian, K., Jin, Z., Hasegawa-Johnson, M., & Mysore, G. J. (2020, May). F0-consistent many-to-many non-parallel voice conversion via conditional autoencoder. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 6284-6288). IEEE.

• the official code and demo which I referred are following:

  • https://github.com/auspicious3000/autovc

  • https://auspicious3000.github.io/autovc-demo/

  • https://auspicious3000.github.io/icassp-2020-demo/

• Additional refer:

  • auspicious3000/autovc#33 (comment)

• torch >= 1.5.0

• tensorboardX >= 2.0

• librosa >= 0.7.2

• matplotlib >= 3.1.3

• Optional for losses flow

  • tensorboard >= 2.2.2

• Currently uploaded code is compatible with the following datasets.
• The O mark to the left of the dataset name is the dataset actually used in the uploaded result.

Before proceeding, please set the pattern, inference, and checkpoint paths in 'Hyper_Parameter.yaml' according to your environment.

• Sound

  • Setting basic sound parameters.

• Content_Encoder

  • Setting the parameters of content encoder.

• Style_Encoder

  • Setting the parameters of style encoder.
  • Encoder is a pre-trained speaker embedding model.
    • https://github.com/CODEJIN/Speaker_Embedding_Torch
  • All parameters must be matched to pre-trained speaker embedding.

• Decoder

  • Setting the parameters of decoder.

• Postnet

  • Setting the parameters of convolution postnet.

• WaveNet

  • Setting the parameters of Vocoder.
  • This implementation uses a pre-trained Parallel WaveGAN model.
    • https://github.com/CODEJIN/PWGAN_Torch
  • If checkpoint path is null, model does not exports wav files.
  • If checkpoint path is not null, all parameters must be matched to pre-trained Parallel WaveGAN model.

• Train

  • Setting the parameters of training.
  • When the number of speaekrs in your train dataset is small, I recommend to increase the Train_Pattern/Accumulated_Dataset_Epoch.

• Inference_Path

  • Setting the inference path

• Checkpoint_Path

  • Setting the checkpoint path

• Log_Path

  • Setting the tensorboard log path

• Device

  • Setting which GPU device is used in multi-GPU enviornment.
  • Or, if using only CPU, please set '-1'.

python Pattern_Generate.py [parameters]

At least, one or more of datasets must be used.

• -vctk
  • Set the path of VCTK. VCTK's patterns are generated.
• -vc1
  • Set the path of VoxCeleb1. VoxCeleb1's patterns are generated.
• -vc2
  • Set the path of VoxCeleb2. VoxCeleb2's patterns are generated.
• -libri
  • Set the path of LibriTTS. LibriTTS's patterns are generated.
• -cmua
  • Set the path of CMU Arctic. CMU Arctic's patterns are generated.
• -vc1t
  • Set the path of VoxCeleb1 testset. VoxCeleb1's patterns are generated for an evaluation.
• -mw
  • The number of threads used to create the pattern

python Train.py -s <int>

• -s <int>
  • The resume step parameter.
  • Default is 0.
  • When this parameter is 0, model try to find the latest checkpoint in checkpoint path.

• Current training....

• Current training....