Shanghai-based AI lab StepFun has launched StepAudio 2.5 Realtime, a cutting-edge voice AI model designed for real-time audio processing without text conversion. The model, capable of handling both Chinese and English, is tailored for conversational voice agents, particularly in extended roleplay scenarios. StepAudio 2.5 boasts advanced paralinguistic awareness, detecting non-verbal cues such as speech rate and emotional tone, and maintains persona stability through roleplay-specific reinforcement learning. StepFun's internal benchmarks show StepAudio outperforming existing models in paralinguistic comprehension and conversational quality. The company, founded by former Microsoft veteran Jiang Daxin, positions StepAudio as a competitor to OpenAI's voice mode, claiming superior performance. The model is now live, with the initial persona "Xiao Yue" available for public interaction, and developers can create custom personas via the API. This innovation could significantly impact crypto and Web3 applications, enhancing social dApps, metaverse interactions, and voice-enabled trading assistants.