StepFun 13B parameter open-source end-to-end multimodal audio LLM featuring studio-grade cloning and emotion control.
Weights and inference scripts fully open-sourced on GitHub and Hugging Face under permissive terms
Step-Audio is an open-source 13-billion-parameter multimodal audio foundation model engineered by pioneering AI lab StepFun. Bypassing traditional cascaded voice pipelines, Step-Audio adopts a unified Transformer architecture directly mapping speech latents to linguistic semantics and expressive audio synthesis. It excels at zero-shot voice replication, dialect preservation, and fine-grained emotional modulation—allowing developers to direct whisper, excitement, or melancholy via simple natural language prompts.