inclusionAI/Ming-UniAudio

**Summary for Developer Directory:** Built on MingTok-Audio, a unified continuous speech tokenizer using VAE with causal Transformer architecture, Ming-UniAudio integrates semantic and acoustic features to enable a single LLM backbone for both speech understanding and generation tasks. It introduces instruction-guided free-form speech editing without timestamp requirements, powered by a Diffusion Head for synthesis quality. Models are available on Hugging Face and ModelScope with ASR/TTS fine-tuning recipes and the first open-source free-form audio editing benchmark.

435 stars.