Skip to content

full tensorflow implementation of the paper: StarGAN-VC: Non-parallel many-to-many voice conversion with star generative adversarial networks https://arxiv.org/abs/1806.02169

License

Notifications You must be signed in to change notification settings

marlon-br/StarGAN-Voice-Conversion

 
 

Repository files navigation

StarGAN Voice Conversion

This is a tensorflow implementation of the paper: StarGAN-VC: Non-parallel many-to-many voice conversion with star generative adversarial networks.

The converted voice examples are in converted directory

Dependencies

  • Python 3.6 (or higher)
  • tensorflow 1.8
  • librosa
  • pyworld
  • tensorboard
  • scikit-learn

NOTE:According to some feedbacks, we recommend to use tensorflow version 1.8 exactly. (Tensorflow 1.11 generate nonsense results)

Usage

Download dataset

Download the vcc 2016 dataset to the current directory and create train directory and test directory.

python download.py --datasets vcc2016 --train_dir ./data/fourspeakers --test_dir ./data/fourspeakers_test

For simplicity use:
python download.py 

The downloaded zip files are extracted to ./data/vcc2016_training and ./data/evaluation_all.

  1. training set: In the experiment, we choose four speakers from ./data/vcc2016_training. We move the corresponding folder(eg. SF1,SF2,TM1,TM2 ) to ./data/fourspeakers.
  2. testing set In the experiment, we choose four speakers from ./data/evaluation_all. We move the corresponding folder(eg. SF1,SF2,TM1,TM2 ) to ./data/fourspeakers_test.

The data directory now looks like this:

data
├── fourspeakers  (training set)
│   ├── SF1
│   ├── SF2
│   ├── TM1
│   └── TM2
├── fourspeakers_test (testing set)
│   ├── SF1
│   ├── SF2
│   ├── TM1
│   └── TM2
├── vcc2016_training (vcc 2016 training set)
│   ├── ...
├── evaluation_all (vcc 2016 evaluation set, we use it as testing set)
│   ├── ...

Preprocess dataset

Extract features (mcep, f0, ap) from each speech clip. The features are stored as npy files. We also calculate the statistical characteristics for each speaker.

python preprocess.py --input_dir ./data/fourspeakers --output_dir ./data/processed --ispad True

For simplicity use:
python preprocess.py

This process may take a few minutes !

Note that test set doesn’t need preprocess.

Train

Read npy files from processed_dir to train model and raw wav files from test_wav_dir to randomly generate some samples using the model during training.

python train.py --processed_dir ./data/processed --test_wav_dir ./data/fourspeakers_test

For simplicity use:
python train.py

Convert

Restore model from model_dir, convert source_speaker’s speech to target_speaker’s speech. The results are strored in ./converted_voices

python convert.py --model_dir ./your_model_dir  --source_speaker SF1 --target_speaker TM1

Summary

The network structure shown as follows:

Snip20181102_2

Note: Our implementation follows the original paper’s network structure, while pytorch StarGAN-VC code‘network is different from the paper as it’s classifier shares the Discriminator’s weights. Both ways generate good converted speeches.

Reference

CycleGAN-VC code

pytorch StarGAN-VC code

StarGAN code

StarGAN-VC paper

StarGAN paper

CycleGAN paper


If you feel this repo is good, please star !

Your encouragement is my biggest motivation!

About

full tensorflow implementation of the paper: StarGAN-VC: Non-parallel many-to-many voice conversion with star generative adversarial networks https://arxiv.org/abs/1806.02169

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%