dimalishudi/tts

DLA HW3: Text To Speech via FastSpeech

Homework 3 (TTS)

Implements FastSpeech1/2 models via PyTorch. Wandb report

Project Navigation

• hw_tts/ is main module of this project; it directly implements FastSpeech and FastSpeech2.
• requirements.txt lists all python requirements.
• preprocess_data.py implements data preprocessing.
• setup.sh script installs needed requirements, downloads data and preprocesses it.
• train.py and synthesis.py scripts are used to train model and synthesise audio respectively.
• launcher.ipynb notebook shows possible ways to run scripts for this project.
• figs/ directory contains some figures for report.

Requirements and Set Up

To install all requirements, download needed data and preprocess it simply run:

./setup.sh

Model Training and Audio Synthesis

To train model run

python3 ./train.py -c ./hw_tts/configs/fs2_train_config.json

To synthesise audio run

python3 ./synthesis.py -c ./hw_tts/configs/fs2_synthesis_config.json -w ./fs2_saved/checkpoint_19999.pth.tar

All scripts menthioned above are also shown in launcher.ipynb notebook.

References

• FastSpeech 2: Fast and High-Quality End-to-End Text to Speech, Y. Ren, et al.
• xcmyz's FastSpeech implementation
• ming024's FastSpeech2 implementation