Résumé:
Deepfake audio technology poses a growing threat to information authenticity and in-
tegrity. This thesis provides a systematic investigation of different Machine Learning
(ML) methods for detecting deepfake in Arabic speech. Firstly, a novel dataset of real
and synthetic Arabic audio speech was created. Then, various ML methods were evaluated
for their ability to discriminate between genuine and synthesized speech. Finally, a new
Arabic deepfake speech framework is proposed, including handcrafted feature extraction
and classification. Feature importance analysis revealed key acoustic and prosodic cues
that contribute to the detection process, where the XGBoost classifier emerged as the
most effective. Experimental results demonstrated the robustness and the high accuracy
of our proposed framework for Arabic deepfake speech detection compared to state-of-the-
art methods. This research establishes a benchmark for Arabic deepfake audio detection
and contributes to the ongoing efforts to combat the harmful effects of this technology.
