#large-margin data

Step-Audio-EditX is an open-source 3B audio LLM that edits speech at the token level, enabling iterative, controllable changes to emotion, style, and paralinguistic cues through large-margin synthetic data and token-level PPO.