A model with 7 billion parameters that speaks the Emirati dialect, a transcription model for six languages and an Arabic text reader: the Emirati institute is focusing on languages that are underserved by large language models.
The Technology Innovation Institute (TII) in Abu Dhabi presented on October 7 three new models in the Falcon family focused on Arabic. Falcon-Emirati, with 7 billion parameters, is designed to converse in the dialect of the Emirates; it obtains 84.83% on Alyah, a test bed dedicated to native Emirati Arabic, and becomes accessible via the Falcon Chat platform.
Compact yet accurate transcription
Falcon-ASR, with 1.6 billion parameters, converts speech to text in Modern Standard Arabic, Emirati Arabic, English, French, Spanish and Portuguese. On an internal test in the UAE dialect, it outperforms larger systems, including a multimodal model with 30 billion parameters. It can timestamp each word, making it easier to search through long records. The third model, Falcon-OCR-Arabic, extracts Arabic text from scanned documents.
The TII is aimed at concrete uses: multilingual customer service, transcription of meetings and interviews, subtitling, accessibility and digitization of documents. The announcement is part of the UAE's strategy of sovereign models, with particular attention to dialects for which transcribed training data remains scarce.
Read the full article on the original source
Go to the source →