StepFun debuts StepAudio 3 voice model family, tops real-time benchmarks

StepFun released StepAudio 3, a family of five audio models covering real-time voice conversation, speech-to-text transcription, text-to-speech generation, sound-effect generation, and music. The real-time conversation model is built to handle interruptions, reason while still speaking, and call external tools mid-conversation rather than only converting speech to text and back.

StepFun's own thread claims the real-time model ranks #1 on Artificial Analysis's Conversational Dynamics and Speech Reasoning categories, at 98.9% and 99.7% respectively, and that its transcription model reaches 1.7% word error rate. Artificial Analysis's own leaderboard corroborates the transcription result, placing StepAudio 3's ASR at the top of its speech-to-text accuracy index, effectively tied with Alibaba's Fun-Realtime-ASR-preview at the same 1.7% rate; the conversational-dynamics and speech-reasoning figures come only from StepFun's own announcement and were not independently checked here.