Step-Audio-EditX opens its training stack alongside the model

StepFun updated Step-Audio-EditX, a speech-editing model, and released the training code behind it: SFT, DPO and GRPO, three different post-training methods for adjusting a model's behavior after its base training, plus a vLLM-based inference setup the company says speeds up both training and generation. Releasing that code, rather than only the trained weights, is what lets an outside team retune the model's emotion and pacing control against their own recordings instead of accepting StepFun's shipped voice behavior.

StepFun's own description of the update claims fidelity gains in emotion and style control, finer control over speech rate, and new paralinguistic tags, markers for non-word vocal cues like laughter or breathing. Those improvement claims come directly from StepFun's release notes rather than an independent evaluation.