我的语音合成最新进展见 ParallelTTS。

Go ParallelTTS for my latest work of TTS.

Tacotron-2 的 PyTorch 实现，见 Tacotron2-PyTorch。

PyTorch implementation of Tacotron-2, See Tacotron2-PyTorch.

tacotron-2-mandarin

Tensorflow implementation of DeepMind's Tacotron-2. A deep neural network architecture described in this paper: Natural TTS synthesis by conditioning Wavenet on MEL spectogram predictions

Repo Structure

tacotron-2-mandarin-griffin-lim
|--- datasets
|--- logs-Tacotron
     |--- eval-dir
     |--- plots
     |--- taco_pretrained
     |--- wavs
|--- papers
|--- prepare
|--- tacotron
     |--- models
     |--- utils
|--- tacotron_output
     |--- eval
     |--- logs-eval
          |--- plots
          |--- wavs
|--- training_data
     |--- audio
     |--- linear
     |--- mels

Samples

There are some synthesis samples here.

Pretrained

you can get pretrained model here.

Quick Start

OS: Ubuntu 16.04

Step (0) - Git clone repository

git clone https://github.com/atomicoo/tacotron2-mandarin.git
cd tacotron-2-mandarin-griffin-lim/

Step (1) - Install dependencies

Install Python 3 (python-3.5.5 for me)
Install TensorFlow (tensorflow-1.10.0 for me)
Install other dependencies
```
pip install -r requirements.txt
```

Step (2) - Prepare dataset

Download dataset BIAOBEI or THCHS-30

After that, your doc tree should be:

tacotron-2-mandarin-griffin-lim
|--- ...
|--- BZNSYP
     |--- ProsodyLabeling
          |--- 000001-010000.txt
     |--- Wave
|--- ...

Prepare dataset (default is BIAOBEI)
```
python prepare_dataset.py
```
If preparing THCHS-30, you can use parameter --dataset=THCHS-30.

After that, you can get a folder BIAOBEI as follow:
```
tacotron-2-mandarin-griffin-lim
|--- ...
|--- BIAOBEI
     |--- biaobei_48000
|--- ...
```

Preprocess dataset (default is BIAOBEI)

python preprocess.py

If prrprocessing THCHS-30, you can use parameter --dataset=THCHS-30.

After that, you can get a folder training_data as follow:

tacotron-2-mandarin-griffin-lim
|--- ...
|--- training_data
     |--- audio
     |--- linear
     |--- mels
     |--- train.txt
|--- ...

Step (3) - Train tacotron model

python train.py

More parameters, please see train.py.

After that, you can get a folder logs-Tacotron as follow:

tacotron-2-mandarin-griffin-lim
|--- ...
|--- logs-Tacotron
     |--- eval-dir
     |--- plots
     |--- taco_pretrained
     |--- wavs
|--- ...

Step (4) - Synthesize audio

python synthesize.py

More parameters, please see synthesize.py.

After that, you can get a folder tacotron_output as follow:

tacotron-2-mandarin-griffin-lim
|--- ...
|--- tacotron_output
     |--- eval
     |--- logs-eval
          |--- plots
          |--- wavs
|--- ...

References & Resources

Rayhane-mamah/Tacotron-2

atomicoo / tacotron2-mandarin