Alibaba Launches Qwen-Audio-3.0-ASR-Flash for Enhanced Industry-Specific Speech Recognition
Alibaba has launched Qwen-Audio-3.0-ASR-Flash, a speech recognition model optimized for context consistency, industry-specific terminology, and hotword customization. The model is available on the Alibaba Cloud Bailian platform in three versions, including offline file transcription and real-time streaming. This update addresses a common limitation in speech-to-text systems by significantly improving the recognition of rare and domain-specific terms, achieving a 95.36% accuracy rate in medical scenarios. It also ranked first on the Artificial Analysis benchmark with a low 1.7% character error rate, making it highly viable for enterprise applications like meeting minutes and customer service. The model features speech polishing capabilities to directly output structured text and supports audio inputs up to 5 minutes for the Flash version. The training involved building a high-quality vocabulary database covering fields such as medicine, IT programming, finance, and public figures.
## BACKGROUND
Automatic Speech Recognition (ASR) systems often struggle with specialized jargon, acronyms, or uncommon names, which is typically addressed using "hotword customization" to bias the model toward specific terms. Alibaba Cloud's Bailian platform (also known as Model Studio) is a comprehensive development platform that allows users to access, fine-tune, and deploy various foundation models like the Qwen series.