yl4579/StyleTTS2

Combines latent diffusion for style generation with pre-trained speech language models (WavLM) as adversarial discriminators, enabling reference-free style synthesis and differentiable duration modeling for end-to-end training. Provides pre-trained components including multilingual text aligners, pitch extractors, and PL-BERT models, supporting single and multi-speaker training on datasets like LJSpeech, VCTK, and LibriTTS with zero-shot speaker adaptation capabilities. Built on PyTorch with configuration-driven training pipelines and Hugging Face integration for inference.

6,205 stars. No commits in the last 6 months.