StepFun's StepAudio 3 ASR Reaches 1.7% Word Error Rate, Tops Speech Recognition Leaderboard

StepFun has released StepAudio 3 ASR, a speech-to-text model that now leads the Artificial Analysis Word Error Rate (AA-WER) benchmark at 1.7%, matching Alibaba's Fun-Realtime-ASR and representing a three-percentage-point improvement over the company's previous StepAudio 2.5 (4.7%). The model is available via the StepFun API for non-streaming speech recognition applications.