StepFun’s Voice AI Topped Every Benchmark. It Also Hears Your Sighs

Shanghai-based AI lab StepFun released StepAudio 2.5 Realtime this week. It’s an end-to-end real-time voice model—audio goes in, audio comes out, no text conversion in the middle. It supports Chinese and English and, based on benchmarks, it seems to be pretty good.

The lab is best known for building text LLMs that outperform much larger systems. Step 3.5 Flash, a 196-billion-parameter model, topped four reasoning benchmarks earlier this year against trillion-parameter rivals. (Parameters are what give an AI model its breadth of knowledge, and more generally translates to better performance.)

The voice work follows the same playbook, and wants to make roleplay cool, especially in longer sessions.

The character problem

AI persona systems have a specific failure mode: OOC, or out-of-character behavior—the model drifts from its assigned personality under adversarial pressure. It’s embarrassingly common, and is a flaw that exists in all AI models by design. They just forget stuff the more you interact with them.

StepFun says it solved this with roleplay-specific RLHF—reinforcement learning from human feedback applied specifically to persona stability, not just general quality. The training data starts from over 10,000 human-authored persona seeds, algorithmically expanded into a million-scale feature matrix.

The idea: enough variety in the training data that even weird, long-tail conversations don’t knock the model off its character.

The more technically interesting claim is paralinguistic comprehension—the model reads non-verbal acoustic cues like vocal speed, emotional tone, and age from the audio itself, before it formulates a response.

On the paralinguistic comprehension benchmark—an objective test measuring acoustic feature perception like emotion and speaking rate, scored 0–100—StepAudio hit 82.18. GPT Realtime 1.5 scored 80.46, Gemini Live hit 58.05, and DouBao Realtime came in at 16.09.

The human evaluation benchmark—real users talking to the model via a mobile app, scored by human raters on a 0–100 scale—came in at 80.41 for StepAudio, against 68.01 for GPT Realtime 1.5 and 67.16 for Gemini Live. General dialogue quality, tested objectively via API on the same 0–100 scale, landed at 86.36 versus GPT’s 81.60.

These are StepFun’s own benchmarks. Make of that what you will. But the margins on paralinguistics and spoken question and answer sessions are large enough that they’re hard to dismiss.

The StepFun Context

StepFun was founded in April 2023 by Jiang Daxin, who spent 16 years at Microsoft running projects like Bing, Cortana, and Azure cognitive services. It’s one of China’s so-called AI Tiger startups and has raised roughly $1.7 billion to date.

OpenAI’s advanced voice mode launched in late 2024 and set the benchmark everyone else is chasing. StepFun is now benchmarking against it directly—and claiming a win.

The launch includes a flagship AI persona called Xiao Yue, which StepFun describes as a “soul-level companion” designed to feel like texting a friend, not querying software. Opinions, catchphrases, emotional limits—fully configurable.

Developers can build their own personas via the API. Full documentation is at platform.stepfun.com, and the model is live now.