StepAudio 2.5 Realtime reads tone and holds a persona

Per StepFun's post, StepAudio 2.5 Realtime is a live voice model that StepFun says reads paralinguistic cues, tone, pace, pauses, and things like a mid-sentence half-laugh, and lets a developer define a persona, personality, backstory, language style, through the API rather than choosing from a small set of preset voices. StepFun says it offers more than 10,000 native personas plus five ready-made presets, and that the model was tuned with reinforcement learning from human feedback to keep holding a chosen character under adversarial roleplay tests.

This is a different product from StepAudio 2.5 TTS, which StepFun released earlier in the same line and which is directed after the fact by writing plain-language instructions about emotion and pacing into a text-to-speech request, with zero-shot voice cloning that keeps a cloned voice's tone and emotional control separately adjustable (StepAudio 2.5 TTS). Realtime, by contrast, is built for a live spoken exchange rather than generating narration from text.