From the source
Ant Ling introduced Ming-UniAudio, an end-to-end unified speech language model with a single LLM backbone for both understanding and generation tasks, enhanced with a Diffusion Head for high-fidelity speech synthesis.
It uses MingTok-Audio, a unified continuous speech tokenizer based on a VAE framework with a causal Transformer architecture, claimed to be the first continuous speech tokenizer to integrate semantic and acoustic features.
The release also includes Ming-Freeform-Audio-Edit, an instruction-guided free-form speech editing framework, and Audio-Edit-Benchmark, an open-source evaluation set for free-form speech editing tasks.
