Classification of Libyan Dialects Using Wav2Vec2 XLSR-53 and Deep Learning Techniques
Year: 2026
Type: preprint
Abstract: This study addresses the classification of Libyan Arabic dialects using deep learning techniques and speech representations extracted from the pre-trained Wav2Vec2 XLSR-53 model. It focuses on distinguishing Western, Eastern, and Southern dialectal groups, based on a set of audio recordings collected from publicly available sources. The study explores two classification approaches: the first relies on frozen speech representations with a Random Forest classifier, and the second relies on fine-tuning the Wav2Vec2 model for the dialect classification task. It also employs exploratory analysis techniques, including Principal Component Analysis (PCA) and Self-Organizing Maps (SOM), to examine the distribution patterns of speech representations and the relationships between them. The results provide empirical evidence on the potential of leveraging pre-trained speech representation models for classifying the Libyan dialectal groups under study, and establish a foundation for future research addressing data expansion, improvement of evaluation methods, and investigation of model generalization to different speakers and audio sources. Keywords: Libyan Arabic dialects, Wav2Vec2 XLSR-53, Deep Learning, Random Forest, Speech Representations, Principal Component Analysis. Note: This graduation project was defended at the University of Tripoli, Faculty of Science, Department of Computer Science, on September 22, 2026.
Author: Abdul Karim, Ritaj Mahmoud
Institution: University of Tripoli
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225018258