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

Artificial Analysis, an independent benchmarking firm, ranks StepFun's new StepAudio 3 ASR first on its index for non-streaming speech-to-text, the kind that transcribes a finished recording rather than live audio. The measure is word error rate, the share of words a model transcribes incorrectly. StepAudio 3 ASR scores 1.7%, tied with Alibaba's Fun-Realtime-ASR-preview and down from 4.7% for StepFun's previous model, StepAudio 2.5 ASR. StepFun had already claimed the 1.7% figure when it launched its StepAudio 3 family days earlier; this post is the independent measurement.

How the index is built explains the result. According to the chart in the post, half of the score comes from AA-AgentTalk, Artificial Analysis's own unpublished set of speech directed at voice agents, and a quarter each from recordings of parliamentary proceedings and long corporate earnings calls. StepAudio 3 ASR leads on the voice-agent set at 1.4%, but on the earnings calls it scores 2.8% against 1.8% for Alibaba's model. Its top rank therefore reflects strength on short conversational speech. For long recordings like these earnings calls, the Alibaba model did better on this data.

The price is the other half of the trade. StepAudio 3 ASR costs $0.40 per hour of audio, or $6.67 per 1,000 minutes, the most of the five most accurate models. MAI-Transcribe-2 and Grok Voice Transcribe 2.0, at 2.0% and 2.3% error in the chart, cost $1.67 per 1,000 minutes, and ElevenLabs Scribe v2, at 2.2%, costs $3.67. When StepFun launched StepAudio 2.5 ASR, it listed it at $0.022 per hour, a small fraction of the new model's price. The difference between 1.7% and 2.0% is about three fewer wrong words per thousand. That gain matters most to voice-agent builders whose systems act on each transcribed word, and less to anyone transcribing large volumes where price dominates. It is also not the fastest option: it runs at 88 times real time, against 374 times for MAI-Transcribe-2, though well ahead of Fun-Realtime-ASR-preview at 20 times.