StepFun's StepAudio 3 ASR Tops Speech-to-Text Accuracy Leaderboard

Speech-to-text accuracy is usually measured by word error rate (WER), the share of words a transcription model gets wrong compared with a human transcript. Artificial Analysis reports that StepFun's new StepAudio 3 ASR model now tops its non-streaming Speech-to-Text leaderboard with a 1.7% WER, a sharp improvement over StepFun's own prior model, StepAudio 2.5 ASR, which scored 4.7%, and the first time a StepFun model has led that leaderboard.

That top score is effectively tied with Alibaba's Fun-Realtime-ASR-preview, also at 1.7%. Looking at specific conditions, StepAudio 3 ASR leads on Artificial Analysis's held-out voice-agent dataset at 1.4% WER, but trails on long-form earnings-call audio at 2.8%, where Fun-Realtime-ASR-preview scores 1.8%, meaning its accuracy advantage is not uniform across audio types.

The tradeoff is speed and price: StepAudio 3 ASR transcribes at about 88x real time, slower than several rivals like MAI-Transcribe-2 at 374x, and costs $0.40 per hour of audio, the most expensive of the five most accurate models Artificial Analysis tracks. The reported takeaway is that it buys top-tier accuracy on conversational audio specifically, at a higher per-minute cost than faster competitors.