Skip to content

Latest commit

 

History

History
109 lines (81 loc) · 2.8 KB

README.md

File metadata and controls

109 lines (81 loc) · 2.8 KB
Raw

A Vietnamese TTS

Tacotron + HiFiGAN vocoder for vietnamese datasets.

A synthesized audio clip: clip.wav. A colab notebook: notebook.

Install

git clone https://github.com/NTT123/vietTTS.git
cd vietTTS 
pip3 install -e .

Quick start using pretrained models

bash ./scripts/quick_start.sh

Download InfoRe dataset

bash ./scripts/download_aligned_infore_dataset.sh

Note: this is a denoised and aligned version of the original dataset which is donated by the InfoRe Technology company (see here). You can download the original dataset (InfoRe Technology 1) at here.

Train duration model

python3 -m vietTTS.nat.duration_trainer

Train acoustic model

python3 -m vietTTS.nat.acoustic_trainer

Train HiFiGAN vocoder

We use the original implementation from HiFiGAN authors at https://github.com/jik876/hifi-gan. Use the config file at assets/hifigan/config.json to train your model.

git clone https://github.com/jik876/hifi-gan.git

# create dataset in hifi-gan format
ln -sf `pwd`/train_data hifi-gan/data
cd hifi-gan/data
ls -1 *.TextGrid | sed -e 's/\.TextGrid$//' > files.txt
cd ..
head -n 100 data/files.txt > val_files.txt
tail -n +101 data/files.txt > train_files.txt
rm data/files.txt

# training
python3 train.py \
  --config ../assets/hifigan/config.json \
  --input_wavs_dir=data \
  --input_training_file=train_files.txt \
  --input_validation_file=val_files.txt

Finetune on Ground-Truth Aligned melspectrograms:

cd /path/to/vietTTS # go to vietTTS directory
python3 -m vietTTS.nat.zero_silence_segments -o train_data # zero all [sil, sp, spn] segments
python3 -m vietTTS.nat.gta -o /path/to/hifi-gan/ft_dataset  # create gta melspectrograms at hifi-gan/ft_dataset directory

# turn on finetune
cd /path/to/hifi-gan
python3 train.py \
  --fine_tuning True \
  --config ../assets/hifigan/config.json \
  --input_wavs_dir=data \
  --input_training_file=train_files.txt \
  --input_validation_file=val_files.txt

Then, use the following command to convert pytorch model to haiku format:

cd ..
python3 -m vietTTS.hifigan.convert_torch_model_to_haiku \
  --config-file=assets/hifigan/config.json \
  --checkpoint-file=hifi-gan/cp_hifigan/g_[latest_checkpoint]

Synthesize speech

python3 -m vietTTS.synthesizer \
  --lexicon-file=train_data/lexicon.txt \
  --text="hôm qua em tới trường" \
  --output=clip.wav