StepFun's new voice AI model, StepAudio 2.5 Realtime, surpasses competitors in benchmarks, showcasing advanced persona stability and paralinguistic understanding.
StepAudio 2.5 Realtime Outperforms Major Competitors
StepFun, a Shanghai-based AI lab, has announced the launch of StepAudio 2.5 Realtime, a cutting-edge real-time speech model that excels in voice processing. This innovative model supports both Chinese and English and boasts a fully customizable persona feature.
In benchmark tests conducted in April 2026, StepAudio outperformed notable competitors, including GPT Realtime 1.5 and Gemini Live, across all five evaluated voice AI benchmarks. The model was developed using a comprehensive dataset of one million personas and refined through roleplay-specific reinforcement learning, aimed at mitigating the common issue of character drift during extended interactions.
Addressing Character Drift in AI
Character drift, where AI systems deviate from their established personas, has been a significant challenge in AI interactions. StepFun claims to have effectively tackled this issue by leveraging human feedback in their training process, ensuring that the personas remain stable even in complex conversational scenarios.
StepAudio also features advanced paralinguistic capabilities, allowing it to interpret non-verbal signals such as emotional tone and speaking rates. In the paralinguistic comprehension benchmark, StepAudio achieved an impressive score of 82.18, surpassing competitors like GPT Realtime 1.5, which scored 80.46.
In the human evaluation benchmark, where real users rated the model's performance, StepAudio received a score of 80.41, again outpacing GPT Realtime 1.5 and Gemini Live, which scored 68.01 and 67.16, respectively. The general dialogue quality assessment also favored StepAudio, which scored 86.36 compared to GPT's 81.60.
Background and Future Prospects
Founded in April 2023 by Jiang Daxin, a former Microsoft executive, StepFun has positioned itself as a leading AI startup in China, raising approximately $1.7 billion in funding. The company is now benchmarking its technology against advanced voice solutions from OpenAI, which launched in late 2024.
The flagship persona from StepAudio, named Xiao Yue, is designed to simulate a more human-like interaction, allowing users to customize various attributes for a more personalized experience. Developers can create their own personas using the API, with comprehensive documentation available for integration.
