StepFun Releases StepAudio 2.5 Realtime: An End-to-End Voice Model with Roleplay-Specific RLHF and Paralinguistic Comprehension

StepFun, a Shanghai-based artificial intelligence laboratory, has unveiled StepAudio 2.5 Realtime, a sophisticated end-to-end voice model designed to deliver real-time speech interactions with unprecedented personalization capabilities. Released in May 2026, this advancement represents a significant leap in conversational AI technology, introducing specialized training methodologies and paralinguistic comprehension that distinguish it from existing voice models in the market.

StepAudio 2.5 Realtime operates as a fully end-to-end real-time speech large language model with WebSocket API connectivity, enabling seamless integration across multiple platforms and applications. The model demonstrates multilingual support, handling both Chinese and English with native-level proficiency. Most notably, the system achieved first-place rankings across all five evaluated benchmarks, establishing new performance standards for voice-based AI systems.

The model's architecture incorporates roleplay-specific reinforcement learning from human feedback (RLHF), a training methodology that enables users to customize personas and conversational styles with exceptional granularity. This specialized approach allows the system to maintain consistent character traits, emotional nuances, and linguistic patterns aligned with user specifications.

• Paralinguistic comprehension capabilities enable the model to interpret tone, emotion, and contextual nuance beyond traditional speech recognition
• Fully customizable persona features open new applications in entertainment, customer service, and educational technologies
• Real-time performance eliminates latency issues that have historically limited voice AI adoption
• Cross-language support positions the model competitively in global AI markets
• Advanced RLHF training methodology establishes new benchmarks for conversational AI development
• WebSocket API architecture facilitates enterprise-level integration and deployment

StepFun's release underscores the rapid evolution of conversational AI toward more natural, emotionally intelligent interactions. As businesses increasingly prioritize user experience in voice-enabled applications, technologies like StepAudio 2.5 Realtime address critical gaps in naturalness, customization, and responsiveness. The model's performance achievements validate sophisticated training approaches that balance technical excellence with human-centered design, positioning voice AI as a practical tool for diverse commercial and creative applications rather than an emerging experimental technology.