VITS for Japanese

VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech

Several recent end-to-end text-to-speech (TTS) models enabling single-stage training and parallel sampling have been proposed, but their sample quality does not match that of two-stage TTS systems. In the repository, I will introduce a VITS model for Japanese on pytorch version 2.0.0 that customed from VITS model.

We also provide the pretrained models.

VITS at training	VITS at inference

Pre-requisites

0. Python >= 3.6
1. Clone this repository
2. Install python requirements. Please refer requirements.txt
3. Download datasets
   1. Download and extract the Japanese Speech dataset, then choose basic5000 dataset and move to jp_dataset folder.
4. Run preprocessing if you use your own datasets.

   # Preprocessing (g2p) for your own datasets. Preprocessed phonemes for Japanese dataset have been already provided.
   python preprocess.py --text_index 1 --filelists filelists/jp_audio_text_train_filelist.txt filelists/jp_audio_text_val_filelist.txt filelists/jp_audio_text_test_filelist.txt

Training Example

# JP Speech
python train.py -c configs/jp_base.json -m jp_base

Inference Example

To get pretrained model for Japanese:

sh startup.sh

See vits_apply.ipynb or run streamlit run app.py to see demo on streamlit-share.