Advancing Arabic ASR for Disordered Speech: Fine-Tuning Wav2Vec2 on Egyptian Dysarthric Speech

Abstract

Despite significant advances in Automatic Speech Recognition (ASR), its application to low-resource languages such as Arabic—especially for speakers with speech disorders—remains underdeveloped. This study presents a novel approach to Arabic ASR for disordered speech by fine-tuning a Wav2Vec2 model on a personalized dataset comprising approximately 1,300 utterances from an Egyptian Arabic speaker with speech impairments. Building on the comparative foundation set by Alsohby (2025), which evaluated four state-of-the-art ASR models across general, dysarthric, and accented speech, we extend the analysis through specialized model adaptation. Our methodology encompasses data preprocessing, fine-tuning, and evaluation using Word Error Rate (WER) and Character Error Rate (CER). Results indicate a substantial performance gain, reducing WER from 0.8516 to 0.3736 and CER from 0.5756 to 0.3478. These findings demonstrate the effectiveness of personalized fine-tuning and underscore the critical need for diverse, domain-specific datasets to improve ASR accessibility for Arabic speakers with speech impairments.