Alibaba's new speech recognition model hits 95% accuracy in medical settings

Alibaba released its new speech recognition model Qwen-Audio-3.0-ASR-Flash on July 31, aiming to improve AI's understanding of professional terms. The model is optimized for context consistency, industry-specific word recognition, and custom hotwords, and it can output structured text directly. The team built a high-quality vocabulary from fields like healthcare, IT, stocks, and celebrities, boosting recognition of technical terms. In internal tests, accuracy in medical scenarios exceeded 95.36%. The Qwen-Audio-ASR-Flash series is already used in meeting notes, live captions, educational recordings, and customer service. It once ranked first globally on Artificial Analysis with a 1.7% character error rate. The new model is now available via Alibaba Cloud's Bailian platform, offering three versions: Flash for up to 5-minute recognition, Filetrans for offline transcription, and Streaming for real-time recognition.