Voine/Bert-VITS2-MNN

Implements full offline Chinese TTS inference on Android by combining a distilled 30MB BERT model with MNN-optimized Bert-VITS2 components, using cppjieba and tokenizers-cpp to replace Python dependencies. The pipeline handles text tokenization, G2P, BERT embeddings, and waveform generation entirely on-device without network access. Model conversion chains PyTorch → ONNX → MNN format with 8-bit weight quantization for mobile deployment.

129 stars.