September 22, 2026
Conversation transcription has fewer errors: StepAudio 3 ASR cut them to 1.7%
On September 22, StepAudio 3 ASR scored 1,7% errors on the AA-WER index, tying Alibaba for first place. The model processes uploaded recordings through the StepFun API, with billing based on actual usage.
Artificial Analysis
@artificialanlys
StepFun has released StepAudio 3 ASR, which took 1st place in the AA-WER index among models for non-streaming speech transcription: 1,7% WER. This is notably better than StepAudio 2.5 ASR, which scored 4,7%. StepAudio 3 ASR is StepFun's new speech-to-text model. It is available for non-streaming transcription through the StepFun API and became the first StepFun model to top our non-streaming speech-transcription ranking. Key points ➤ Accuracy: StepAudio 3 ASR scores 1,7% on the AA-WER index, effectively sharing first place with Alibaba Fun-Realtime-ASR-preview, which also scores 1,7%. On AA-AgentTalk, our private dataset of voice-agent dialogues, the model leads with 1,4%, but it trails on long Earnings22 calls: 2,8% versus 1,8% for Fun-Realtime-ASR-preview. ➤ Speed: the model transcribes audio 88 times faster than real time. It trails MAI-Transcribe-2 at 374×, Smallest AI Pulse Pro at 285×, and Grok Voice Transcribe 2.0 at 154×; is roughly on par with ElevenLabs Scribe v2 at 84× and Gemini 3.5 Transcribe at 91×; and outpaces Fun-Realtime-ASR-preview at 20×. ➤ Price: StepAudio 3 ASR costs $0,40 per hour, or $6,67 per 1000 minutes. It is the most expensive of the five most accurate models. MAI-Transcribe-2 and Grok Voice Transcribe 2.0 cost $1,67 per 1000 minutes, while ElevenLabs Scribe v2 costs $3,67, so leader-level conversational-speech accuracy comes at a higher price. More details below ⬇️
· 18.3K views
StepAudio 2.5 ASR previously made 4,7% errors on AA-WER. StepAudio 3 ASR now makes 1,7%, and the index is built from roughly 8 hours of audio, with voice-agent dialogues accounting for half its weight.
At EmpirioLabs, `stepaudio-3-asr-max` is called via `POST /v1/audio/transcriptions`. The service accepts WAV, MP3, OGG, M4A, and PCM.
Speed and price. StepAudio transcribes a recording 87,9 times faster than its duration, nearly matching Gemini 3.5 Transcribe at 91,1×. One thousand minutes cost $6,67 with StepAudio, compared with $0,37 for StepAudio 2.5 ASR.
For non-streaming transcription, the choice now comes down not only to error rate, but also to cost per minute.
