GST-Tacotron-Pytorch

A PyTorch implementation of Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis

model

Update

Add blizzard dataset support.

Requirements

pip3 install -r requirements.txt

File structure

Hyperparameters.py --- contain all hyperparameters
Network.py --- encoder and decoder
Modules.py --- some modules for tacotron
Loss.py --- loss function
Data.py --- load dataset
utils.py --- some util functions for data I/O
Synthesis.py --- generate wav files

How to train

Download a multispeaker dataset
preprocess your data and write your get_XX_data function in Data.py
Adjust hyperparameters in Hyperparameters.py
make a directory named log as follow:

--- log
|    |
|    --- log[log_number]
|
--- code
     |
     --- Tacotron
             |
             --- train.py
             |
             --- Network.py
             |
           ......

run train.py

python3 train.py [log_number] [dataset_size] [start_epoch]

[log_number]: the log directory number
[dataset_size]: int or all
[start_epoch]: which epoch start to train (0 if start from scratch )

for example:
python3 train.py 0 all 0

How to generate wav

rungenerate.py, modify the textin generate.py before running

only support Chinese