Jieyue Xingchen Releases StepAudio 2.5 Realtime: Creating More 'Human-like' AI Voice Dialogue Experiences

Jieyue Xingchen recently released its new real-time voice large model StepAudio 2.5 Realtime, which is now fully online. Developers can access it via the company's open platform. The model aims to create more 'human-like' AI dialogue experiences, with upgrades focused on three key areas: paralinguistic perception, character customization, and dialogue ability.

The core breakthrough lies in paralinguistic processing—interpreting non-verbal expressions like tone, speed, pauses, sighs, and laughs to detect the speaker's emotions (e.g., tiredness from a low voice, irritability from fast speech) and dynamically adjust response strategies to reduce interaction stiffness.

For character customization, users can fine-tune AI roles' traits (personality, background, etc.) via API. It relies on a million-level trait matrix generated from over 10,000 high-quality native personas, optimized with RLHF to avoid 'character collapse' and maintain consistency even in extreme scenarios. Five preset personas are also available.

April 2026 evaluations show the model ranks first in five dimensions: subjective human dialogue score is 80.41 (higher than GPT-Realtime-1.5's 68.01 and Gemini Live's 67.16), and its voice Q&A benchmark score is 1.5 times that of GPT-Realtime-1.5.