fishaudio/Bert-VITS2

Integrates pretrained multilingual BERT encoders for phoneme/semantic representation instead of traditional text encoders, enabling cross-lingual voice synthesis within a VITS2 generative framework. The architecture combines BERT's contextual embeddings with VITS2's flow-based decoder and adversarial training for improved naturalness across multiple languages. Designed for training custom voice models from scratch with preprocessing pipelines exposed via `webui_preprocess.py`.

8,707 stars. Actively maintained with 1 commit in the last 30 days.