StepAudio 2.5 Tops Five Voice AI Benchmark Tests

According to CryptoBriefing, StepFun's StepAudio 2.5 Realtime speech model ranked first in five major benchmark tests in April 2026, achieving a subjective human evaluation score of 80.41, a 10-point improvement over its predecessor. This model employs an end-to-end architecture, supports real-time interaction in both Chinese and English, and incorporates paralinguistic understanding to recognize intonation, emotion, and speech rate. The technical report indicates that it maintains role consistency through role-playing-specific RLHF technology, distinguishing it from traditional speech recognition-language model-speech synthesis pipelines.