Qwen Releases Qwen-Audio-3.0-ASR: Supports Long-Context and 30 Languages

千问大模型 · wechat · 2026-07-31

Alibaba's Qwen has officially released the Qwen-Audio-3.0-ASR-Flash speech recognition model, focusing on solving professional vocabulary recognition and long-audio consistency. It previously ranked first globally on ArtificialAnalysis with a 1.7% Character Error Rate.

The new model introduces four major upgrades:

Long Context Memory: Can reference previously identified content during long audio transcription, automatically updating memory to significantly reduce homophone and terminology errors.
Industry Vocab & Hotwords: Built-in high-quality multi-industry lexicons (medical scenario hit rate exceeds 95.36%) and supports instant hotword customization with over 99% hit rate in most scenarios without false triggers.
One-Step Speech Polish: Directly removes filler words, cleans up stuttering, and handles self-corrections during the recognition phase, outputting structured written text comparable to a two-step 'recognition + LLM polish' approach.
Multilingual Integration: A single model supports 30 languages including Chinese, Japanese, Korean, Southeast Asian, and European languages. The average semantic error rate across seven languages is only 17.09%, outperforming Azure and Gemini 3.0 Flash.

Additionally, the Streaming version for low-latency scenarios reduces the Chinese error rate in complex industrial contexts to 7.8% while maintaining a 300ms latency. The models are now available on Alibaba Cloud's Bailian platform.

Original post →

Qwen Releases Qwen-Audio-3.0-ASR: Supports Long-Context and 30 Languages

More from Models