Emirati dialect-trained AI models are released by Abu Dhabi’s Technology Innovation Institute

The speech recognition model has 1.6 billion parameters, but it was more accurate than a 30-billion-parameter model, according to TII. Abu Dhabi’s Technology Innovation Institute, a state-run research body, released AI models trained on the Emirati Arabic dialect, improving translation, transcription, and voice-enabled services. The Falcon-Emirati model and related tools for speech recognition and extracting Arabic text from images and documents are relatively compact. The speech recognition model has 1.6 billion parameters, but is more accurate than a 30-billion-parameter model, according to TII. AI training in Arabic has lagged behind other languages, partly because dialects vary widely and can be incomprehensible to people who know only the formal Arabic used in news reports and textbooks. This has made some machine translations comically inaccurate, sometimes turning innocuous words into vulgar references in English. TII aims to eliminate those errors by training models on local dialects, idioms, and cultural references, ultimately making AI more useful in the region.