Dual-CycleGAN for NonParallel High-Quality Audio Super-Resolution

This repositry provides official pytorch implementation of Dual-CycleGAN.
Specifically, Dual-CycleGAN enables you to train a high-quality super resolution (SR) model (e.g., 16kHz -> 48kHz) only with low-resolution audio signals of the target domain with high-resolution audio signals of another domain.

Please check the DEMO for more information.

Environment setup

cd DualCycleGAN
pip install -e .

This will install the core library (dual-cyclegan and its submodules) and CLI tools (e.g., dual-cyclegan-train). Please refer to the Parallel WaveGAN repo for more details.

Folder architecture

• egs: The folder for projects.
• egs/tts_16kHz.gt_48kHz: The folder of the "TTS + Ground truth" project example.
• dual_cyclegan: The folder of the source codes.

Run

In this repo, hyperparameters are managed using Hydra.
Hydra provides an easy way to dynamically create a hierarchical configuration by composition and override it through config files and the command line.

Dataset preparation

Make dataset and list files denoting paths to each audio files according to your own dataset (E.g., egs/tts_16kHz.gt_48kHz/data/list/tts_train_16kHz.list). Note that list files for each training/validation/evaluation are required.

Training

# Train a model customizing the hyperparameters as you like
$ dual-cyclegan-train data=tts_16kHz.gt_48kHz model=dual_cyclegan train=dual_cyclegan out_dir=exp/dual_cyclegan

Inference

# Infer with a trained model from a checkpoint file
$ dual-cyclegan-infer data=tts_16kHz.gt_48kHz model=dual_cyclegan out_dir=exp/dual_cyclegan/wav checkpoint_path=exp/dual_cyclegan/checkpoint-600000steps.pkl

Monitor training progress

tensorboard --logdir exp

Details of the list files

List files contain the path to audio files used for training, validation, and evaluation, respectively.
A set of the three files indicates distribution of a single dataset.
Please note that Dual-CycleGAN requires two sets of list files because it is trained on two kinds of datasets.
Please check egs directory for examples of the intended directory structure.

Details of the list files

List files contain the path to audio files used for training, validation, and evaluation, respectively.
A set of the three files indicates distribution of a single dataset.

# An example of the project directory structure
cd DualCycleGAN/egs/tts_16kHz.gt_48kHz/
  |- data
  |    |- list
  |        |- tts_train_no_dev.list
  |        |- tts_dev.list
  |        |- tts_eval.list
  |        |- gt_train_no_dev.list
  |        |- gt_dev.list
  |        |- gt_eval.list
  |- exp
       |- dual_cyclegan

# An example of the list file
path_to_your_own_dataset_dir/audio1.wav
path_to_your_own_dataset_dir/audio2.wav
path_to_your_own_dataset_dir/audio3.wav
...