fatchord/WaveRNN

PyTorch implementation of DeepMind's WaveRNN vocoder with integrated Tacotron text-to-speech synthesis, using mel-spectrogram conditioning and Mixture of Logistics output for efficient neural audio generation. The system decouples acoustic modeling (Tacotron) from vocoding (WaveRNN), allowing independent training pipelines and optional Ground-Truth Alignment (GTA) datasets to improve synthesis quality. Includes pretrained models on LJSpeech and supports both batched and unbatched inference modes.

2,179 stars. No commits in the last 6 months.