Jieyue Xingchen Releases StepAudio 2.5 ASR Model: 400% Speed Boost, 90% Price Cut

Jieyue Xingchen recently launched its new automatic speech recognition model StepAudio 2.5 ASR, with core breakthroughs in applying large language model inference acceleration technology to speech recognition. The model's inference speed increased by 400%, latency reduced by 60%, inference cost dropped by 80%, and its pricing is only 0.15 yuan per hour, about one-tenth of the previous generation product.

StepAudio 2.5 ASR uses an ASR+MTP-5 deep fusion architecture, transplanting the multi-token prediction technology from large language models to break the efficiency bottleneck of traditional autoregressive mechanisms. It also reuses the 32K context window capability of large language models, supporting end-to-end processing of up to 30 minutes of continuous audio without segmentation, solving the problem of context fragmentation in long audio transcription.

The model is suitable for scenarios like meeting transcription, voice interaction, media content processing, etc. It is now fully available on Jieyue Xingchen's open platform and Step Plan, and developers can experience it via the official website.