Editorial illustration for StepFun launches StepAudio 2.5 Realtime, evaluated via mobile app raters
StepFun launches StepAudio 2.5 Realtime, evaluated via...
The numbers tell a compelling story. StepAudio 2.5 Realtime didn’t just edge past competitors; it swept every benchmark dimension, from subjective mobile chat scores to paralinguistic nuance, with commanding leads. The evaluation itself was anything but synthetic: human raters scored real-time conversations through actual mobile app interactions, a testing framework that prizes ecological validity over sterile lab conditions.
StepFun’s latest end-to-end speech LLM achieves an 80.41 on subjective human evaluation, an 86.36 on general dialogue, and an 82.18 on paralinguistic comprehension, a technical differentiator that lets the model hear not just words, but the tilt of a voice, the pace of a sigh, the flicker of emotion in a sentence. Persona-specific RLHF and million-scale data augmentation lock in character consistency across turns, while WebSocket API access at `wss://api.stepfun.com/v1/realtime` makes deployment immediate. The result is a voice model that doesn’t just listen, it understands how you sound.
Starting from 10,000+ high-quality natively authored personas, StepFun applied algorithmic augmentation to build a million-scale persona feature matrix. This was combined with millions of real-world conversational samples for training. The intent is generalization — specifically, stable performance on difficult, long-tail conversational topics.
StepAudio 2.5 Realtime’s 80.41 human evaluation score wasn’t earned in a sterile lab. It was fought for in the messy, unpredictable flow of real mobile conversations, rated by actual people. That’s the difference.
The model doesn’t just hear words; it reads the room. It catches the sarcasm in a clipped tone, the urgency in a faster pace, the hesitation before a pause. Paralinguistic comprehension is not a gimmick here, it’s the engine.
And with persona-specific RLHF and million-scale data augmentation, every interaction stays consistent without turning robotic. Shanghai-based StepFun has delivered an end-to-end speech LLM that doesn’t just talk back. It listens.
The API is live at `wss://api.stepfun.com/v1/realtime`, model string `step-2.5-realtime`. The demo is waiting. The model card is open.
The next layer of voice interaction just got real.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Anthropic may keep supplying Claude to NSA despite Pentagon risk flag
From Archives Claude Code auto‑creates AI scaling algorithms; new control allocates compute
Common Questions Answered
What evaluation method did StepFun use to test StepAudio 2.5 Realtime?
StepFun evaluated StepAudio 2.5 Realtime using human raters who scored real-time conversations through actual mobile app interactions, rather than synthetic lab conditions. This testing framework prioritizes ecological validity by measuring how the model performs in messy, unpredictable real-world mobile conversations with actual users.
What was StepAudio 2.5 Realtime's human evaluation score and how did it compare to competitors?
StepAudio 2.5 Realtime achieved an 80.41 human evaluation score and swept every benchmark dimension, achieving commanding leads over competitors. The model demonstrated superior performance across subjective mobile chat scores and paralinguistic nuance evaluation metrics.
How does StepAudio 2.5 Realtime handle paralinguistic comprehension?
StepAudio 2.5 Realtime uses paralinguistic comprehension as its core engine to detect nuances beyond words, such as sarcasm in clipped tones, urgency in faster pacing, and hesitation before pauses. The model achieves this capability through persona-specific RLHF and million-scale data augmentation techniques.
What makes StepAudio 2.5 Realtime an end-to-end speech LLM?
StepAudio 2.5 Realtime is an end-to-end speech LLM that processes real-time audio conversations directly without requiring intermediate transcription steps. This architecture allows the model to maintain paralinguistic information and contextual nuances throughout the entire conversation flow.
Further Reading
- Papers with Code - Latest NLP Research — Papers with Code
- Hugging Face Daily Papers — Hugging Face
- ArXiv CS.CL (Computation and Language) — ArXiv
