Stepwise Launch StepAudio2.5ASR: introduces the MTP mechanism to accelerate, and a 32K window eliminates transcription fragments

robot
Abstract generation in progress
ME News reports that on April 24 (UTC+8), according to Dongcha Beating monitoring, Jieyue Xingchen released a new generation automatic speech recognition model, StepAudio 2.5 ASR, which has now been fully launched on its open platform.
This version is the first to introduce the multi-token prediction (MTP) technology of large language models into the field of speech recognition. While significantly improving inference speed, it reuses the large model's 32K context window, breaking the limitation of traditional long audio transcription that requires slicing and splicing.
Traditional speech recognition is limited by the autoregressive mechanism, which requires outputting tokens one by one.
StepAudio 2.5 ASR has ported the same ASR+MTP-5 deep fusion architecture as Step 3.5 Flash, predicting multiple candidate tokens at once and verifying them in parallel.
The official says that this architecture increases model inference throughput by 400%, reduces latency by 60%, and reduces inference cost by 80%, with a peak inference rate of 500 tokens/s.
To address the context fragmentation problem caused by the widely used "slice-transcribe-splice" approach in the industry (e.g., forgetting the beginning background when transcribing the latter part), the new model directly reuses the 32K context window, supporting end-to-end single reading of complete audio up to 30 minutes long.
In a 30-minute full-load input test, the model did not show any accuracy degradation over time.
Its comprehensive error rate on 10 authoritative open-source test sets in Chinese and English, such as LibriSpeech, is lower than that of competitors.
(Source: BlockBeats)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pinned