A Comparative Study of Arabic Dialect Identification Using Machine Learning and Transfer Learning 

Khalid Mansour Email 

Faculty of Computer Studies, Arab Open University, A'ali, 732, Bahrain

Abstract

Arabic Dialect Identification (ADI) is a challenging task in speech processing, especially under limited data conditions. In this work, we present a comprehensive experimental study of ADI using the publicly available portion of the Casablanca dataset, covering eight Arabic dialects: Algerian, Egyptian, Emirati, Jordanian, Mauritanian, Moroccan, Palestinian, and Yemeni. We systematically evaluate the impact of four complementary strategies: traditional machine learning with MFCC acoustic features, statistical feature engineering, data augmentation via Gaussian noise injection, SpecAugment, and Deep Convolutional Generative Adversarial Networks (DCGANs), and transfer learning through fine-tuning of the Wav2Vec2 XLSR-53 model. Our results demonstrate a clear performance hierarchy: statistical feature engineering significantly improves traditional ML models, DCGAN-based data synthesis modestly outperforms simpler augmentation methods by generating realistic dialect-specific feature maps, and pre-trained transformer models deliver the strongest performance, reaching 92.34% accuracy and substantially outperforming all other approaches. A dedicated ablation further confirms that freezing the convolutional feature encoder during fine-tuning matches or exceeds full fine-tuning while updating substantially fewer parameters. Across all experiments, Jordanian Arabic was consistently the most accurately identified dialect, while Yemeni Arabic remained the most challenging for traditional approaches. Our findings provide practical insights into which techniques are most beneficial for Arabic dialect identification under limited data conditions.

A Comparative Study of Arabic Dialect Identification Using Machine Learning and Transfer Learning