Tongyi Launches Fun-ASR1.5 with Enhanced Dialect Recognition

ME News reports that on April 20 (UTC+8), according to monitoring by Beating, Tongyi Lab released the speech recognition model Fun-ASR 1.5 on April 20, with its API now available on Alibaba Cloud’s Bailian and an online demo accessible via the ModelScope community. The official statement indicates that this version uses a single model to cover 30 languages, seven major Chinese dialect groups, and over 20 regional accents, eliminating the need for separate models per dialect. Internal evaluations by Tongyi show that the character error rate in typical dialect scenarios has decreased by 56.2% compared to the previous version, with five dialects achieving over 90% accuracy and 15 exceeding 80%. Special optimization has also been applied to classical poetry recognition, with an internal character-level accuracy of 97%. All figures are based on Tongyi’s internal testing and are not third-party benchmarks. The long-tail challenge of Chinese dialect speech recognition is now integrated into a unified, commercially viable system. For use cases such as educational live streaming, local government hotlines, and interview transcription, users no longer need to deploy multiple separate recognition pipelines for different regional accents, simplifying deployment. (Source: BlockBeats)