What it is
Open-source foundation model that unifies speech generation and instruction-guided editing from text and optional audio context, plus a distilled 4-step fast variant. Instead of separate task-specific speech models, one model handles generation, editing, enhancement, and separation via natural-language instructions plus optional audio context; it couples MLLM semantic conditioning with shared audio latents and a unified flow generator.
Key results
- AuK-Flash gives 4.5× wall-clock speedup vs 32-NFE teacher
- AuK-Flash runs in 4 steps without classifier-free guidance
- Trained on 3.03B instruction-audio pairs and 1.95M hours
A single instruction-conditioned speech model can cover broad generation and editing tasks without giving up strong generation quality.