OpenBMB/VoxCPM

Employs an end-to-end diffusion autoregressive architecture built on the MiniCPM-4 backbone that directly generates continuous speech representations, avoiding discrete tokenization bottlenecks. Achieves semantic-acoustic decoupling through hierarchical language modeling and FSQ constraints, enabling streaming synthesis at RTF ~0.15 on consumer GPUs. Supports both full-parameter and LoRA fine-tuning via Hugging Face and ModelScope, with optional speech enhancement and ASR integration through ZipEnhancer and SenseVoice.

6,143 stars and 17,933 monthly downloads. Actively maintained with 2 commits in the last 30 days. Available on PyPI.