Chinese mandarin text to speech based on Fastspeech2 and Unet

This is a modification and adpation of fastspeech2 to mandrin(普通话）. Many modifications to the origin paper, including:

Use UNet instead of postnet (1d conv). Unet is good at recovering spect details and much easier to train than original postnet
Added hanzi(汉字，chinese character) embedding. It's harder for human being to read pinyin, but easier to read chinese character. Also this makes it more end-to-end.
Removed pitch and energy embedding, and also the corresponding prediction network. This makes its much easier to train, especially for my gtx1060 card. I will try bringing them back if I have time (and hardware resources)
Use only waveglow in synth, as it's much better than melgan and griffin-lim.
subtracted the mel-mean for (seems much) easier prediction.
Changed the loss weight to mel_postnet_loss x 1.0 + d_loss x 0.01 + mel_loss x 0.1
Used linear duration scale instead of log, and subtracted the duration_mean in training.

Audio samples

New audios for aishell3 added. Checkpoints and training scripts for multispeaker will be ready soon.

Audio samples for biaobei and aishell3 can also be found in this page

Dependencies

All experiments were done under ubuntu16.04 + python3.7 + torch 1.7.1. Other env probably works too.

torch for training and inference
librosa and ffmpeg for basic audio processing
pypinyin用于转换汉字为拼音
jieba 用于分词
perf_logger用于写训练日志

First clone the project

git clone https://github.com/ranchlai/mandarin-tts.git

If too slow, try

git clone https://hub.fastgit.org/ranchlai/mandarin-tts.git

To install all dependencies, run


sudo apt-get install ffmpeg
pip3 install -r requirements.txt

Synthesize

python synthesize.py --input="您的电话余额不足，请及时充值"

(see this issue for help)

or put all text in input.txt, then

python synthesize.py --input="./input.txt"

Checkpoints and waveglow should be downloaded at 1st run. You will see some files in ./checkpoint, and ./waveglow

In case it fails, download the checkpoint manully here

Model architecture

Training

(under testing)

Currently I am use baker dataset(标贝）, which can be downloaded from baker。 The dataset is for non-commercial purpose only, and so is the pretrained model.

I have processed the data for this experiment. You can also try

python3 preprocess_pinyin.py 
python3 preprocess_hanzi.py

to generate required aligments, mels, vocab for pinyin and hanzi for training. Everythin should be ready under the directory './data/'(you can change the directory in hparams.py) before training.