Oven logo

Oven

Published

PhaseAug: A Differentiable Augmentation for Speech Synthesis to Simulate One-to-Many Mapping

pip install phaseaug

Package Downloads

Weekly DownloadsMonthly Downloads

Project URLs

Requires Python

>=3

PhaseAug

PhaseAug: A Differentiable Augmentation for Speech Synthesis to Simulate One-to-Many Mapping
Junhyeok Lee, Seungu Han, Hyunjae Cho, Wonbin Jung @ MINDsLab Inc., SNU, KAIST

arXiv GitHub Repo stars githubio

Abstract : Previous generative adversarial network (GAN)-based neural vocoders are trained to reconstruct the exact ground truth waveform from the paired mel-spectrogram and do not consider the one-to-many relationship of speech synthesis. This conventional training causes overfitting for both the discriminators and the generator, leading to the periodicity artifacts in the generated audio signal. In this work, we present PhaseAug, the first differentiable augmentation for speech synthesis that rotates the phase of each frequency bin to simulate one-to-many mapping. With our proposed method, we outperform baselines without any architecture modification. Code and audio samples will be available at https://github.com/mindslab-ai/phaseaug.

Accepted to ICASSP 2023

phasor

TODO

  • PyTorch 2.0 is released, need to modify STFT and iSTFT for complex support (solved at 1.0.0)
  • Arxiv updated
  • Errata in paper will be fixed. Section 2.5 in paper, transition band half-width 0.06-> 0.012.
  • Section 2.5, mention about multiplyinng rotation matrix to "the left side of F(x)" will be added. -> transpose m,k to reduce ambiguity
  • Upload PhaseAug to pypi.
  • Upload VITS+PhaseAug sampels at demo page.
  • Refactoring codes for packaging.

Use PhaseAug

  • Install alias-free-torch==0.0.6 and phaseaug
pip install alias-free-torch==0.0.6 phaseaug 
  • Insert PhaseAug in your code, check train.py as a example.
from phaseaug.phaseaug import PhaseAug
...
# define phaseaug
aug = PhaseAug()
...
# discriminator update phase
aug_y, aug_y_g = aug.forward_sync(y, y_g_hat.detach())
y_df_hat_r, y_df_hat_g, _, _ = mpd(aug_y, aug_y_g)
y_ds_hat_r, y_ds_hat_g, _, _ = msd(aug_y, aug_y_g)
...
# generator update phase
aug_y, aug_y_g = aug.forward_sync(y, y_g_hat)
y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = mpd(aug_y, aug_y_g)
y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = msd(aug_y, aug_y_g)
  • If you are applying torch.cuda.amp.autocast like VITS, you need to wrap PhaseAug with autocast(enabled=False) to prevent ComplexHalf issue.
from torch.cuda.amp import autocast
...
with autocast(enabled=True)
    # wrapping PhaseAug with autocast(enabled=False)
    with autocast(enabled=False)
        aug_y, aug_y_g = aug.forward_sync(y, y_g_hat)
    # usually net_d parts are inside of autocast(enabled=True)
    y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = net_d(aug_y, aug_y_g)

Requirements

docker build -t=phaseaug --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) --build-arg USER_NAME=$USER

Training

  1. Clone this repository and copy python files to hifi-gan folder
git clone --recursive https://github.com/mindslab-ai/phaseaug
cp ./phaseaug/*.py ./phaseaug/hifi-gan/
cd ./phaseaug/hifi-gan
  • optional: MelGAN generator
cp ./phaseaug/config_v1_melgan.json ./phaseaug/hifi-gan/
  • Change generator to MelGAN generator at train.py
# import MelGanGenerator as Generator at [train.py](./train.py)
#from models import Generator # remove original import Generator
from models import MelGANGenerator as Generator
  1. Modify dataset path at train.py
     parser.add_argument('--input_wavs_dir',
                         default='path/LJSpeech-1.1/wavs_22k')
     parser.add_argument('--input_mels_dir',
                         default='path/LJSpeech-1.1/wavs_22k')
  1. Run train.py
python train.py --config config_v1.json --aug --filter --data_ratio {0.01/0.1/1.} --name phaseaug_hifigan
python train.py --config config_v1_melgan.json --aug --filter --data_ratio {0.01/0.1/1.} --name phaseaug_melgan

References

This implementation uses code from following repositories:

This README and the webpage for the audio samples are inspired by:

Citation & Contact

If this repostory useful for yout research, please consider citing!

@inproceedings{phaseaug,
  author={Lee, Junhyeok and Han, Seungu and Cho, Hyunjae and Jung, Wonbin},
  title={{PhaseAug: A Differentiable Augmentation for Speech Synthesis to Simulate One-to-Many Mapping}},
  journal = {arXiv preprint arXiv:2211.04610},
  year=2022,
}

Bibtex will be updated after ICASSP 2023.

If you have a question or any kind of inquiries, please contact Junhyeok Lee at [email protected]