TPGST reimplementation with pytorch

Paper: PREDICTING EXPRESSIVE SPEAKING STYLE FROM TEXT IN END-TO-END SPEECH SYNTHESIS

Prerequisite

• python 3.7
• pytorch 1.3
• librosa, scipy, tqdm, tensorboardX

Dataset

• KSS, Korean female single speaker speech dataset.

Samples

• Post

Usage

1. Download the above dataset and modify the path in config.py. And then run the below command.

   python prepro.py
   

2. The model needs to train 100k+ steps

   python train.py <gpu_id>
   

3. After training, you can synthesize some speech from text.

   python synthesize.py <gpu_id> <model_path>
   

4. To listen your samples, you may need mel2wav vocoder. I didn't include vocoder in this repo.

Notes

• I think the difference between baseline Tacotron and TPGST is small on KSS dataset.
• I will be doing more experiminets soon.