TTS homework repository of HSE DLA course. The goal of the project is to implement FastSpeech 2 model and train it on LJ-Speech dataset.
To install necessary python packages run the command:
pip install -r requirements.txt
Download all needed resources (data, checkpoints & inference examples) with
python3 bin/download.py
If you use Yandex DataSphere, specify the config
python3 bin/download.py -c datasphere
One may use datasphere.ipynb notebook that contains all necessary commands to reproduce the results of the project.
Generate pitch profiles required for training with
python3 bin/preprocess_pitch.py [-c datasphere]
Once the resources are ready, start the training with
python3 train.py
or
python3 train.py -c datasphere
for DataSphere g1.1 configuration.
Use
python3 inference.py final_model [-c datasphere]
to generate audio files with the following configurations:
- usual generated audio
- audio with +20%/-20% for pitch/speed/energy
- audio with +20/-20% for pitch, speed and energy together
The audio is stored at resources/final_model.
Run python3 some_script.py -h for help.