ASR Models for Traditional Emirati Arabic: Challenges, Adaptations, and Performance Evaluation
Alblooki, Maha Mohammad Abdulla Abdulkarim · MBZUAI iRep · 2025
The rapid advancement of Automatic Speech Recognition (ASR) technology has significantly improved human-computer interaction across various languages and even dialects; however, traditional dialects, particularly the pure Emirati dialect, remain under-represented in modern ASR systems. This research addresses this gap by creating the first traditional Emirati dataset and developing a dialect-specific ASR model capable of accurately transcribing traditional Emirati Arabic, a spoken variety with distinct phonetic, morphological, and syntactic characteristics. This study adopts a data-driven approach, beginning with the compilation of an Emirati dialect speech corpus, sourced from heritage radio broadcasts and literary resources. To effectively model this dialect, the research explores multiple ASR architectures of pre-trained, self-supervised learning models, like Wav2Vec2, XLS-R, Whisper, and MMS, which are finetuned afterwards to assess their adaptability to traditional Emirati Arabic speech, performance across different linguistic variations, and effectiveness in reducing Word Error Rate (WER) and Character Error Rate (CER) compared to baseline models. Through a rigorous evaluation of average WER and CER of the baseline models and the fine-tuned ones, the study demonstrates the effectiveness of fine-tuned, self-supervised models in improving recognition accuracy for dialectal speech. The results indicate that using domain-specific data and linguistic adaptations significantly enhances ASR performance, paving the way for more robust and inclusive Arabic dialectal speech technologies.