Classification of Libyan Dialects Using Wav2Vec2 XLSR-53 and Deep Learning Techniques

Abstract Automatic dialect classification of Arabic speech remains challenging because of substantial linguistic variation among regional dialects and the limited availability of annotated speech resources. This study investigates the classification of three major regional groups of Libyan Arabic: Western, Eastern, and Southern. A dataset comprising 3,509 speech clips was compiled from publicly available sources, including online videos, podcasts, radio broadcasts, and recordings of natural speech. The study investigates pretrained Wav2Vec2 XLSR-53 speech representations through two supervised classification approaches. First, a Random Forest classifier was trained on fixed 1,024-dimensional representations extracted from the pretrained model, without fine-tuning the speech encoder. This baseline experiment achieved an accuracy of 85.33% using an 80:20 stratified train-test split. Second, the pretrained Wav2Vec2 XLSR-53 model was fine-tuned for the three-class dialect classification task. The audio recordings were resampled to 16 kHz and standardized to fixed-length segments of 10 seconds. Time stretching and temporal shifting were applied dynamically during training as data augmentation techniques. The fine-tuned model achieved an accuracy of 97.72% on a held-out test set of 527 clips, with macro-averaged precision, recall, and F1-score of approximately 98%. In addition, Principal Component Analysis (PCA) and Self-Organizing Maps (SOM) were employed independently to explore the organization of speech representations extracted from the pretrained model. These exploratory analyses provide a qualitative view of the distribution and overlap of representations associated with the three dialect groups. The Random Forest and fine-tuned Wav2Vec2 experiments used different data partitions; therefore, their accuracy values are reported as results from separate experimental settings rather than as a strictly controlled direct comparison. This study contributes an empirical investigation of pretrained speech representations for the classification of three underrepresented regional varieties of Libyan Arabic. It also provides complementary supervised and exploratory analyses that may support further research on Libyan Arabic speech technologies. Keywords: Libyan Arabic Dialect Identification, Wav2Vec2 XLSR-53, Random Forest, Speech Representations, Principal Component Analysis, Self-Organizing Maps. Note: This graduation project was defended at the University of Tripoli, Faculty of Science, Department of Computer Science, on September 22, 2026.