StepAudio 2.5 ASR
StepAudio 2.5 ASR is a 4B speech recognition model that uses multi-token prediction. It has file, synchronous, SSE and live WebSocket interfaces.
Match the interface to the recording
File submission suits a completed recording. The stream interface accepts audio chunks while speech is still in progress. The same model family therefore serves different application timing needs.StepFun stepaudio-2.5-asr guide
Handle partial text as partial
Practical guidance. For live captions, keep provisional text separate from the final transcript. A later chunk can change the interpretation of a word. Use timestamps to link the final text to the recording rather than relying on arrival time.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API |
| Model identifiers | stepaudio-2.5-asrstepaudio-2.5-asr-stream |
| Serving variants | The stream suffix selects bidirectional WebSocket recognition. It is grouped here with the same ASR release.StepFun stepaudio-2.5-asr guide |
| File features | The asynchronous file interface supports speaker separation and separate stereo-channel recognition.StepFun stepaudio-2.5-asr guide |
| Live timing | The streaming interface can return word-level timestamps.StepFun bidirectional speech recognition |
Related models
- Step ASR 1.1
StepFun · Step ASR
A speech-recognition model with separate file-transcription and live-streaming variants.
Launch date not confirmedDeprecated API; check access detailsAPI - StepAudio 2 ASR Pro
StepFun · StepAudio 2
StepAudio 2 ASR Pro is the 32B recognition model in StepFun's audio catalog.
Launch date not confirmedAPI - StepAudio 3 ASR Max
StepFun · StepAudio 3
A speech-recognition model combining acoustic modeling with language-model context for Chinese, English and mixed-language recordings.
Launch date not confirmedAPI · Provider product