Friday 17 October 2025 - 11am Speaker: Hanan Aldarmaki (Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI)).Title: Arabic-Centric Speech Processing: Diglossia, Code-Switching, and Other QuirksAbstract: Research on speech and language processing is increasingly moving towards large-scale multilingual and multimodal models. While these developments have led to impressive general-purpose systems, they often overlook the intricacies of specific target languages, each with its own linguistic quirks that can challenge model assumptions and design. Arabic exemplifies this challenge: a language characterized by diglossia, dialectal variations, frequent code-switching and spelling variations. This talk summarizes recent findings on Automatic Speech Processing (ASR) through empirical studies on Arabic-centric pre-training, multi-dialectal fine-tuning, and multilingual pre-training, analyzing their impact on ASR performance across various dialects and conditions. The discussion highlights both the limitations of current models and the value of targeted strategies for robust speech processing systems. Oct 17 2025 11.00 - 12.00 Friday 17 October 2025 - 11am Speaker: Hanan Aldarmaki (Mohamed Bin Zayed University of Artificial Intelligence) Note unusual location: Bayes Centre, G.03
Friday 17 October 2025 - 11am Speaker: Hanan Aldarmaki (Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI)).Title: Arabic-Centric Speech Processing: Diglossia, Code-Switching, and Other QuirksAbstract: Research on speech and language processing is increasingly moving towards large-scale multilingual and multimodal models. While these developments have led to impressive general-purpose systems, they often overlook the intricacies of specific target languages, each with its own linguistic quirks that can challenge model assumptions and design. Arabic exemplifies this challenge: a language characterized by diglossia, dialectal variations, frequent code-switching and spelling variations. This talk summarizes recent findings on Automatic Speech Processing (ASR) through empirical studies on Arabic-centric pre-training, multi-dialectal fine-tuning, and multilingual pre-training, analyzing their impact on ASR performance across various dialects and conditions. The discussion highlights both the limitations of current models and the value of targeted strategies for robust speech processing systems. Oct 17 2025 11.00 - 12.00 Friday 17 October 2025 - 11am Speaker: Hanan Aldarmaki (Mohamed Bin Zayed University of Artificial Intelligence) Note unusual location: Bayes Centre, G.03
Oct 17 2025 11.00 - 12.00 Friday 17 October 2025 - 11am Speaker: Hanan Aldarmaki (Mohamed Bin Zayed University of Artificial Intelligence)
