StepFun Releases StepAudio 3 Series, Tops Global Voice AI Benchmarks

On September 15, 2026, StepFun officially released the StepAudio 3 series of voice models, including StepAudio 3 Realtime, StepAudio 3 ASR, StepAudio 3 TTS, StepAudio 3 Gen, and StepAudio 3 Music. All five models are now available on the StepFun open platform.

According to the company, the series covers real-time voice interaction, full audio content generation, text-to-speech, complex speech understanding, and music creation. Several models topped the Artificial Analysis leaderboards: StepAudio 3 Realtime ranked first globally in the Conversational Dynamics benchmark with a composite score of 98.9%, and also ranked first in Speech Reasoning with a 99.7% accuracy rate. StepAudio 3 ASR achieved a word error rate of 1.7% in the non-streaming speech recognition accuracy ranking, tying for first place globally.

StepAudio 3 Realtime supports native full-duplex real-time conversation, with parallel reasoning and speech generation, as well as asynchronous tool calls and long-running tasks. StepAudio 3 ASR combines large language model context understanding to handle Chinese, English, dialects, code-switching, and professional terminology. StepAudio 3 TTS is designed for real-time interaction and can reproduce laughter, hesitation, and self-corrections. StepAudio 3 Gen can generate voices, sound effects, ambient audio, and background music from natural language descriptions. StepAudio 3 Music supports both generation from scratch and multi-round interactive creation based on ABC notation.