PyTorch implementation of text-to-speech system Fastspeech2. Implemented as a home assignment on deep learning in audio course.

Install requirements by

pip install -r requirements.txt

Before training you should do preprocessing (besides you already have preprocessed data). To do so you should firstly compute statistics by running followig command

cd preprocessing

python3 get_stats.py

Than carefully record gained statistics in appropriate fields of file config/fastspeech2.yaml. And than run command

python3 get_target_pitch_energy.py

Note, that it can take a significant amount of time.

Also note: steps above are describing a situation, when you already got dataset and situated it by path data/LJSpeed-1.1. If it's not the case, than run script by path scripts/download_dataset.sh.

After all preprocessing stuff is done, you may run training by command

cd ..

python3 train.py

All the model settings and training process setups are controllable from the file config/fastspeech2.yaml.

Firstly you should set the path to your pretrained model in field val.model_path in file config/fastspeech2.yaml. My pretrained checkpoint you may get by running scripts/download_checkpoint.sh. After that setup your preferable speed, pitch in energy is neighbour fields. Than set path to file with text you want to synthesize in a field val.val_texts. Than run

python3 eval.py

Generated records will be placed in directory results.