StepFun released its StepAudio 3 lineup on September 15, shipping five models at once — StepAudio 3 Realtime, ASR, TTS, Gen and Music — all live on the StepFun open platform.
According to rankings from third-party evaluator Artificial Analysis, Realtime scored 98.9% on the Conversational Dynamics leaderboard, ranking first; on the same firm's Speech Reasoning leaderboard, its accuracy came in at 99.7%, also first. ASR tied for first on the non-streaming speech recognition accuracy leaderboard with a 1.7% word error rate.
Five models, five jobs
The division of labor breaks down roughly like this: Realtime handles full-duplex real-time dialogue, reasoning directly on audio rather than converting speech to text first; ASR wires the large model's semantic understanding into the recognition pipeline, and StepFun says it can handle dialects, mixed Chinese-English speech and domain-specific vocabulary; TTS produces human-level synthesis that reproduces pauses and tone alongside correct pronunciation; Gen puts voice, sound effects, ambient sound and background music behind a single generation interface; Music supports multi-turn interactive composition and lets users control melody directly using ABC notation.
The most technically demanding of the five is Realtime's “native speech reasoning.” The conventional pipeline converts speech to text, reasons over the text, then converts back to speech — three stages stitched together that lose tone, stress and emotion at the very first step. Skipping the transcription step gives the model a shot at understanding whether “you did such a great job” is genuine praise or sarcasm.
The release said nothing about whether any of the models will be open-sourced or how pricing will work.
How to read that 1.7%
The 1.7% word error rate is a non-streaming result, and that qualifier matters. Non-streaming means the model gets the full audio clip before transcribing it, so it can look at context both before and after any given moment; streaming requires producing text as it listens, seeing only what has already passed, and its error rate typically runs noticeably higher. Comparing numbers across the two modes side by side is meaningless.
The phrase “tied for first” is itself informative. It means the top spot isn't held alone — at least one other model lands on the same decimal point, which shows that in non-streaming recognition, the gap between leaders has narrowed to the third decimal place.
The real differentiation is more likely to sit outside the leaderboard: how many dialects are covered, how cleanly the model switches during code-mixed speech, and whether domain-specific terminology needs extra configuration. StepFun mentions all three, but without comparable numbers.
What it means for domestic teams
Voice models have been coming out of China at a steady clip this year. Tencent's Hunyuan open-sourced AuK, Xiaomi open-sourced a multi-speaker recognition model, and Qwen's voice model gained tool-calling. The difference is the route: the open-source players release weights that anyone can modify; StepFun's five models this time sit directly on its own open platform, called via API — a lower barrier for teams that don't want to build their own inference clusters, at the cost of running on someone else's servers.
For teams building voice products, the decision to adopt usually comes down to three other things: how low Realtime's end-to-end latency can go in milliseconds, whether TTS voices can be customized, and whether the sound effects Gen produces carry commercial-use rights. None of the three is addressed in the release; teams will have to wait for the open platform's documentation.
Sources: StepFun official release, CocoLoop, Artificial Analysis leaderboards, ITHome; the five model names, the Conversational Dynamics and Speech Reasoning scoring methodology, and the non-streaming word error rate are based on the official release and leaderboard pages.