Alibaba Tongyi Lab Launches Fun-ASR1.5 Voice Recognition Large Model, Supporting Multilingual Dialects and Cross-Language Switching

Alibaba Tongyi Lab recently launched the Fun-ASR1.5 voice recognition large model, which supports 30 languages, seven Chinese dialect systems, and over 20 local accents, as well as accurate transcription of ancient poetry recitations. In cross-language switching scenarios, it can automatically recognize and switch without preset language tags, ensuring transcription accuracy.

The average character error rate of Fun-ASR1.5 has decreased by 56.2% compared to the previous version. It can restore dialect-specific vocabulary and achieve a 97% character-level accuracy in ancient poetry recognition. Additionally, the model can automatically insert punctuation and convert spoken numbers, dates, etc., into standard formats, reducing manual proofreading costs. Currently, the model is available for API services on the Alibaba Cloud Bailian platform and open for experience on the ModelScope community.