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

Download the above dataset and modify the path in config.py. And then run the below command.
```
python prepro.py
```
The model needs to train 100k+ steps
```
python train.py <gpu_id>
```
After training, you can synthesize some speech from text.
```
python synthesize.py <gpu_id> <model_path>
```
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.