Unofficial PyTorch implementation of Kaneko et al.'s MaskCycleGAN-VC (2021) for non-parallel voice conversion.
MaskCycleGAN-VC is the state of the art method for non-parallel voice conversion using CycleGAN. It is trained using a novel auxiliary task of filling in frames (FIF) by applying a temporal mask to the input Mel-spectrogram. It demonstrates marked improvements over prior models such as CycleGAN-VC (2018), CycleGAN-VC2 (2019), and CycleGAN-VC3 (2020).
Figure1: MaskCycleGAN-VC Training
Figure2: MaskCycleGAN-VC Generator Architecture
Figure3: MaskCycleGAN-VC PatchGAN Discriminator Architecture
Paper: https://arxiv.org/pdf/2102.12841.pdf
Repository Contributors: Claire Pajot, Hikaru Hotta, Sofian Zalouk
Clone the repository.
git clone [email protected]:GANtastic3/MaskCycleGAN-VC.git
cd MaskCycleGAN-VC
Create the conda environment.
conda env create -f environment.yml
conda activate MaskCycleGAN-VC
The authors of the paper used the dataset from the Spoke task of Voice Conversion Challenge 2018 (VCC2018). This is a dataset of non-parallel utterances from 6 male and 6 female speakers. Each speaker utters approximately 80 sentences.
Download the dataset from the command line.
wget --no-check-certificate https://datashare.ed.ac.uk/bitstream/handle/10283/3061/vcc2018_database_training.zip?sequence=2&isAllowed=y
wget --no-check-certificate https://datashare.ed.ac.uk/bitstream/handle/10283/3061/vcc2018_database_evaluation.zip?sequence=3&isAllowed=y
wget --no-check-certificate https://datashare.ed.ac.uk/bitstream/handle/10283/3061/vcc2018_database_reference.zip?sequence=5&isAllowed=y
Unzip the dataset file.
mkdir vcc2018
apt-get install unzip
unzip vcc2018_database_training.zip?sequence=2 -d vcc2018/
unzip vcc2018_database_evaluation.zip?sequence=3 -d vcc2018/
unzip vcc2018_database_reference.zip?sequence=5 -d vcc2018/
mv -v vcc2018/vcc2018_reference/* vcc2018/vcc2018_evaluation
rm -rf vcc2018/vcc2018_reference
To expedite training, we preprocess the dataset by converting waveforms to melspectograms, then save the spectrograms as pickle files <speaker_id>normalized.pickle and normalization statistics (mean, std) as npz files <speaker_id>_norm_stats.npz. We convert waveforms to spectrograms using a melgan vocoder to ensure that you can decode voice converted spectrograms to waveform and listen to your samples during inference.
python data_preprocessing/preprocess_vcc2018.py \
--data_directory vcc2018/vcc2018_training \
--preprocessed_data_directory vcc2018_preprocessed/vcc2018_training \
--speaker_ids VCC2SF1 VCC2SF2 VCC2SF3 VCC2SF4 VCC2SM1 VCC2SM2 VCC2SM3 VCC2SM4 VCC2TF1 VCC2TF2 VCC2TM1 VCC2TM2
python data_preprocessing/preprocess_vcc2018.py \
--data_directory vcc2018/vcc2018_evaluation \
--preprocessed_data_directory vcc2018_preprocessed/vcc2018_evaluation \
--speaker_ids VCC2SF1 VCC2SF2 VCC2SF3 VCC2SF4 VCC2SM1 VCC2SM2 VCC2SM3 VCC2SM4 VCC2TF1 VCC2TF2 VCC2TM1 VCC2TM2
Train MaskCycleGAN-VC to convert between <speaker_A_id> and <speaker_B_id>. You should start to get excellent results after only several hundred epochs.
python -W ignore::UserWarning -m mask_cyclegan_vc.train \
--name mask_cyclegan_vc_<speaker_id_A>_<speaker_id_B> \
--seed 0 \
--save_dir results/ \
--preprocessed_data_dir vcc2018_preprocessed/vcc2018_training/ \
--speaker_A_id <speaker_A_id> \
--speaker_B_id <speaker_B_id> \
--epochs_per_save 100 \
--epochs_per_plot 10 \
--num_epochs 6172 \
--batch_size 1 \
--lr 5e-4 \
--decay_after 1e4 \
--sample_rate 22050 \
--num_frames 64 \
--max_mask_len 25 \
--gpu_ids 0 \
To continue training from a previous checkpoint in the case that training is suspended, add the argument --continue_train while keeping all others the same. The model saver class will automatically load the most recently saved checkpoint and resume training.
Launch Tensorboard in a separate terminal window.
tensorboard --logdir results/logs
Test your trained MaskCycleGAN-VC by converting between <speaker_A_id> and <speaker_B_id> on the evaluation dataset. Your converted .wav files are stored in results/<name>/converted_audio.
python -W ignore::UserWarning -m mask_cyclegan_vc.test \
--name mask_cyclegan_vc_VCC2SF3_VCC2TF1 \
--save_dir results/ \
--preprocessed_data_dir vcc2018_preprocessed/vcc2018_evaluation \
--gpu_ids 0 \
--speaker_A_id VCC2SF3 \
--speaker_B_id VCC2TF1 \
--ckpt_dir /data1/cycleGAN_VC3/mask_cyclegan_vc_VCC2SF3_VCC2TF1/ckpts \
--load_epoch 500 \
--model_name generator_A2B \
Toggle between A->B and B->A conversion by setting --model_name as either generator_A2B or generator_B2A.
Select the epoch to load your model from by setting --load_epoch.
├── README.md <- Top-level README.
├── environment.yml <- Conda environment
├── .gitignore
├── LICENSE
|
├── args
│ ├── base_arg_parser <- arg parser
│ ├── train_arg_parser <- arg parser for training (inherits base_arg_parser)
│ ├── cycleGAN_train_arg_parser <- arg parser for training MaskCycleGAN-VC (inherits train_arg_parser)
│ ├── cycleGAN_test_arg_parser <- arg parser for testing MaskCycleGAN-VC (inherits base_arg_parser)
│
├── bash_scripts
│ ├── mask_cyclegan_train.sh <- sample script to train MaskCycleGAN-VC
│ ├── mask_cyclegan_test.sh <- sample script to test MaskCycleGAN-VC
│
├── data_preprocessing
│ ├── preprocess_vcc2018.py <- preprocess VCC2018 dataset
│
├── dataset
│ ├── vc_dataset.py <- torch dataset class for MaskCycleGAN-VC
│
├── logger
│ ├── base_logger.sh <- logging to Tensorboard
│ ├── train_logger.sh <- logging to Tensorboard during training (inherits base_logger)
│
├── saver
│ ├── model_saver.py <- saves and loads models
│
├── mask_cyclegan_vc
│ ├── model.py <- defines MaskCycleGAN-VC model architecture
│ ├── train.py <- training script for MaskCycleGAN-VC
│ ├── test.py <- training script for MaskCycleGAN-VC
│ ├── utils.py <- utility functions to train and test MaskCycleGAN-VC
This repository was inspired by jackaduma's implementation of CycleGAN-VC2.