Fact · Speech recognition
2.5percent AA-WER
On Artificial Analysis's speech-to-text leaderboard, StepAudio 3 ASR, StepFun StepFun had an AA-WER of 2.5 percent, its word error rate on the index's three test datasets.
Checked against its source on October 19, 2026. On October 19, 2026 the quote was no longer on the live page, so the source may have changed since.
In the source's own words
“StepAudio 3 ASR, StepFun StepFun 2.5% 79.4 6.67 Details”
Speech to Text (ASR) Providers Leaderboard & Comparison, Artificial Analysis, Artificial Analysis, 2026, Summary of Key Metrics table.
Open the original- What it measures
- Non-streaming AA-WER v2 index: AA-AgentTalk (50 percent), VoxPopuli-Cleaned-AA (25 percent) and Earnings22-Cleaned-AA (25 percent), about 8 hours of audio. Each row is the provider's hosted endpoint for the model.
- What it does not show
- AA-WER is an index, not a single public test set, and one of its datasets is proprietary. The page does not name a language for the index; the v2.0 article builds its text normalizer on Whisper's English normalizer. Figures change as the leaderboard updates.
In the data
A row of Speech Recognition Accuracy: Word Error Rates, as StepAudio 3 ASR, StepFun StepFun: 2.5 percent.
- System
- StepAudio 3 ASR, StepFun StepFun
- Kind
- Cloud service
- Test set
- AA-WER
- Language
- English
- Speech
- Mixed benchmark
- Year
- 2026
- Note
- Language is not named on the leaderboard page; the AA-WER v2.0 article describes a normalizer built on Whisper's English normalizer.