StepAudio 2.5 TTS takes voice direction in plain language

StepFun launched StepAudio 2.5 TTS, a text-to-speech model that the company says takes emotion, pacing, pauses and delivery instructions written in ordinary language rather than through inline tags or preset combinations. Many controllable TTS systems require inserting markup, like a bracketed tag for 'whisper' or 'pause', directly into the script; StepFun's claim is that a plain sentence describing the desired delivery does the same job.

The launch also claims zero-shot voice cloning that keeps timbre and emotional delivery as separately adjustable parameters, so a cloned voice's identity and its emotional performance can reportedly be set independently rather than tied together. StepFun has not published an evaluation showing how reliably the model follows a given plain-language instruction, so the control claim itself remains the company's own description rather than a measured result.