r9y9/deepvoice3_pytorch

Implements guided attention loss and dilated convolutions to improve alignment convergence, supporting both single-speaker and multi-speaker models trained on LJSpeech, VCTK, and JSUT datasets. Uses convolutional sequence-to-sequence architecture with attention mechanisms and language-dependent text processors for English and Japanese. Includes preprocessors compatible with standard TTS datasets and provides preset hyperparameters for reproducible training across different model variants.

1,982 stars. No commits in the last 6 months.