Talking your language: UAE unveils AI tool designed to master Arabic dialects

The Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi has unveiled an advanced Arabic AI tool called Jais 3. This innovative technology aims to master various regional Arabic dialects and adapt to the rapid changes in language. The Jais 3 platform, developed in collaboration with Abu Dhabi's Inception42, utilizes Jais Speech, a sophisticated voice recognition model specifically designed for the Arabic language and its diverse dialects found across the Gulf and North Africa.

The primary objective of Jais 3 is to keep pace with the intricacies of Arabic, a language spoken by over 400 million people worldwide. Haonan Li, a senior research scientist at IFM, emphasized the importance of understanding Arabic in its everyday context, which varies significantly based on geography and generational differences. For instance, Jais 3 must handle the seamless integration of English or French into Arabic speech and revert back to the appropriate language.

Jais Speech, a key component of Jais 3, excels in transcribing Modern Standard Arabic and the dialects spoken by speakers across 19 Arab countries. It meticulously follows speakers as they switch between Arabic and languages like English or French, maintaining the original script for each language. Moreover, Jais Speech identifies the dialect being spoken and can add Arabic vowel marks to transcripts, ensuring accurate representation of the language.

Developing Jais 3 presented several challenges, particularly in achieving comprehensive representation of the various Arabic dialects. The team had to work extensively with limited open data for Gulf dialects, addressing the tendency of open models to favor Modern Standard Arabic over maintaining the requested dialect. Similarly, for Morocco, the available data heavily favored Moroccan Darija, leaving other dialects underrepresented.

To overcome these challenges, the team employed additional data synthesis and filtering techniques to ensure data quality and relevance.

Richard Morton, vice president and executive director of IFM, highlighted the significance of this technology in the context of the UAE's long-term investment in open AI research and its commitment to the UAE National AI Strategy 2031. The launch of Jais 3 marks a step forward in extending the foundation of AI research into speech and agent capabilities, reflecting the UAE's dedication to building the necessary research capacity and talent to drive the country's AI initiatives.

The performance of Jais 3 was evaluated against a suite of 24 Arabic benchmarks by the Institute of Foundation Models and Inception42. The results demonstrated that the largest Jais 3 model outperformed open models on agentic tasks and matched the strongest performance on Arabic understanding. Jais Speech recorded the lowest average word and character error rates among all models, while the seven billion and 32 billion-parameter Jais models achieved the highest average Arabic scores among comparable open models.

Jais 3 not only enhances the practicality of Arabic AI for business applications but also provides robust reasoning and instruction-following capabilities, facilitating multi-step agentic workflows for organizations of varying computing budgets and deployment needs.