StepFun unveils StepAudio 2.5 Realtime with roleplay-tuned speech

StepAudio 2.5 Realtime is an end-to-end speech model tuned with roleplay RLHF and designed to understand tone, pacing, and emotion. StepFun expanded over 10,000 personas to a million examples for training and reported scores of 86.36 in general dialog, 84.80 in vehicle scenarios, and 82.18 in paralinguistic understanding. The model is available via WebSocket, and its arXiv report is published under CC BY 4.0.

Visit original source