88Z AI, 24 April — StepFun has released StepAudio 2.5 ASR, incorporating large language model inference acceleration technology. Built on the ASR+MTP-5 architecture, the model significantly improves speed and reduces costs, achieving SOTA levels in Chinese and English evaluations while supporting transcription of 30-minute long audio clips.
At a glance
- Built on ASR+MTP-5 architecture, boosting inference speed by 400% and cutting costs by 80%
- Supports a 32K context window, enabling single-pass transcription of 30-minute long audio
The update
StepFun has officially launched the new generation automatic speech recognition model, StepAudio 2.5 ASR. The model innovatively introduces large language model inference acceleration technology into the field of speech recognition, utilizing the ASR+MTP-5 architecture. This technological breakthrough increases inference speed by 400%, reduces latency by 60%, achieves a peak processing speed of 500 tokens/s, and significantly cuts costs by 80%. In terms of performance, StepAudio 2.5 ASR has reached SOTA levels across multiple mainstream Chinese and English evaluation benchmarks. The model leverages a 32K context window and possesses powerful long-audio processing capabilities, allowing it to completely transcribe audio content up to 30 minutes in a single pass, balancing efficiency and accuracy.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 阶跃星辰Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.