UAE unveils Jais 3 AI platform for Arabic dialects — The National
The United Arab Emirates has unveiled the Jais 3 artificial intelligence platform, designed to work with regional dialects of Arabic. The development was created by the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi together with Inception42, The National reports.
Jais 3 can answer queries, transcribe conversation recordings, and create documents and presentations. The platform includes the Jais Speech speech recognition model, adapted for Modern Standard Arabic and the dialects of the Gulf countries and North Africa.
Working with dialects
According to the developers, Jais Speech transcribes Modern Standard Arabic and dialects spoken in 19 Arab countries. The model can also track when a speaker switches from Arabic to English or French and retain foreign-language words in their original script. The system identifies the dialect and can add Arabic diacritical marks to the transcription.
More current news is available on the UA.News Telegram channel Telegram.
Combined with language models, Jais Speech enables Jais 3 to accept spoken instructions in Arabic or English and carry out assigned tasks. Haonan Li, a senior research fellow at the Institute of Foundation Models, noted that the team had to work separately on the representation of different regional forms of Arabic, in particular because of the limited amount of open data on Gulf dialects.
Model evaluation
The Institute of Foundation Models and Inception42 evaluated the models using a set of 24 Arabic-language benchmarks, which they released together with the models. According to the developers, the largest version of Jais 3 outperforms other open models in agentic tasks and demonstrates results on par with the strongest models in Arabic language understanding.
The developers also stated that Jais Speech showed the lowest average word and character error rates among the models compared. The 7-billion- and 32-billion-parameter versions, according to their data, received the highest average scores for Arabic among open models of comparable size.
