PPSpeech: Phrase based Parallel End-to-End TTS System
Pytorch implementation of PPSpeech: Phrase based Parallel End-to-End TTS System.
Requirements :
All code written in Python 3.6.2 .
- Install Pytorch
Before installing pytorch please check your Cuda version by running following command :
nvcc --version
pip install torch torchvision
- Installing other requirements :
pip install -r requirements.txt
- To use Tensorboard install
tensorboard version 1.14.0seperatly with supportedtensorflow (1.14.0)
Note:
- In the paper author break a single sentence into phrases by predicting intonation phrase boundaries(L3) using an expanded CRF supporting dynamic features.
- But in this repo for sake of simplicity I divide sentence into phrases by randomly grouping the words together, which definitely not a true prosodic boundaries, which ultimately hurt the quality for text to speech.But it's don't bother me as I code this repo for just experimentation.
- For better quality use some smart/AI based Phase Boundry detection algo as author used in paper.
Pre-processing
python preprocessing.py -d path_of_wavs --config configs\default.yaml
Training
python train.py -o checkpoints -l logs --name "first" --config configs\default.yaml
Inference
python inference.py -c "checkpoints\first\checkpoint_first_32000.pyt" -r "LJ002-0321.npy" --text put_your_text_here --config "configs\default.yaml" --name wave_file_name --mode 1