One-shot Voice Conversion by Separating Speaker and Content Representations with Instance Normalization

This is the official implementation of the paper One-shot Voice Conversion by Separating Speaker and Content Representations with Instance Normalization. By separately learning speaker and content representations, we can achieve one-shot VC by only one utterance from source speaker and one utterace from target speaker. You can found the demo webpage here, and download the pretrain model from here and the coresponding normalization parameters for inference from here.

Dependencies

• python 3.6+
• pytorch 1.0.1
• numpy 1.16.0
• librosa 0.6.3
• SoundFile 0.10.2
• tensorboardX

Differences from the paper

The implementations are a little different from the paper, which I found them useful to stablize training process or improve audio quality. However, the experiments requires human evaluation, we only update the code but not updating the paper. The differences are listed below:

• Not to apply dropout to the speaker encoder and content encoder.
• Normalization placed at pre-activation position.
• Use the original KL-divergence loss for VAE rather than unit variance version.
• Use KL annealing, and the weight will increase to 1.

Preprocess

We provide the preprocess script for two datasets: VCTK and LibriTTS. The download links are below.

• CSTR VCTK Corpus
• LibriTTS Corpus

The experiments in the paper is done on VCTK.

The preprocess code is at preprocess/. The configuation for preprocessing is at preprocess/libri.config and preprocess/vctk.config. Depends on which dataset you used. where:

• segment_size is the segment size for training. Default: 128
• data_dir is the directory to put preprocessed files.
• raw_data_dir is the directory to put the raw data. Like LibriTTS/ or VCTK-Corpus/.
• n_out_speakers is the number of speakers for testing. Default: 20.
• test_prop is the proportion for validation utterances. Default: 0.1
• training_samples is the number of sampled segments for training (we sample it in the preprocess stage). Default: 10000000.
• testing_samples is the number of sampled segments for testing. Default: 10000.
• n_utt_attr is the number of utterances to compute mean and standard deviation for normalization. Default: 5000.
• train_set: only for LibriTTS. The subset used for training. Default: train-clean-100.
• test_set: only for LibriTTS. The subset used for testing. Default: dev-clean.

Once you edited the config file, you can run preprocess_vctk.sh or preprocess_libri.sh to preprocess the dataset.
Also, you can change the feature extraction config in preprocess/tacotron/hyperparams.py

Training

The default arguments can be found in train.sh. The usage of each arguments are listed below.

• -c: the path of config file, the default hyper-parameters can be found at config.yaml.
• -iters: train the model with how many iterations. default: 200000
• -summary_steps: record training loss every n steps.
• -t: the tag for tensorboard.
• -train_set: the data file for training (train if the file is train.pkl). Default: train
• -train_index_file: the name of training index file. Default: train_samples_128.json
• -data_dir: the directory for processed data.
• -store_model_path: the path to store the model.

Inference

You can use inference.py to inference.

• -c: the path of config file.
• -m: the path of model checkpoint.
• -a: the attribute file for normalization ad denormalization.
• -s: the path of source file (.wav).
• -t: the path of target file (.wav).
• -o: the path of output converted file (.wav).

Reference

Please cite our paper if you find this repository 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}
}

Contact

If you have any question about the paper or the code, feel free to email me at [email protected].