Tacotron for research on Chinese speech synthesis and Taiwanese speech synthesis from Chinese input text sequence with different granularities

Train the Tacotron speech synthesis model to synthesize Chinese or Taiwanese speech conditioned on Chinese input text sequence with different granularities

Prerequisites

1. Python packages:

   • python 3.4 or higher
   • tensorflow r1.8 or higher
   • numpy
   • librosa
   • scipy
   • tqdm
   • matplotlib
   • zhon
   • pypinyin
   • bopomofo

2. Clone this repository:

   https://github.com/HappyBall/tacotron.git

Dataset

1. Lecture record Chinese dataset Download

2. DaAiSermon Taiwanese dataset Download

Usage

After downloading the correspond dataset, you can directly train Tacotron model to synthesize Chinese or Taiwanese speech.

Training

Set up the correct path of the dataset and other hyperparameters in hyperparams.py.

Run: python train.py --keep_train False

Parameter --keep_train determines either start a new training or continue training with the existed model which the path should be correctly set up in hyperparams.py.

Synthesis

1. Set up the correct path of the existed model in hyperparams.py.

2. Add Chinese input sequences you want to synthesize into test_sentences.txt.

Run: python synthesize.py

Evaluate

Set up the output file path in evaluate.py.

Run: python evaluate.py

The program will automatically calculate the mean square error between the mel scale spectrogram of synthesized speech and the ground truth then output as a text file.

Hyperparameters of the hyperparams.py

--data: the path of the data directory which contains the wav files
--prepro_path: the path of the preprocessed data directory
--test_data: the path of the text file which contains input text sequences to synthesize speech
--logdir: the path of the directory to save or load models
--logfile: the path of the training log file
--sampledir: the path of the directory to save speech files when synthesizing
--input_mode: mode of the input granularity (word, bopomofo, pinyin, phoneme)
--withtone: input with tone or not, only bopomofo or pinyin input mode use this hyperparameter (True, False)
--n_iter: iteration number of Griffin Lim algorithm
--guided_attention: use guided attention or not when training (True, False)
--schedule_prob: probability of schedule sampling using the ground truth as input
--lr: initial learning rate

Files in this project

data_load.py: data loader for training data and testing data
evaluate.py: calculate the mean square error between the mel scale spectrogram of synthesized speech and the ground truth from testing data
graph.py: define model graph
hyperparams.py: set up training hyperparameters and directory for saving models
modules.py: define modules like gru, convolution layers ...etc
networks.py: define networks like encoder, decoder ...etc
prepro.py: preprocess data
synthesize.py: synthesize speech conditioned on input text sequences in the test sentence file
train.py: train models