Minale AshagrieOromia Chala2026-10-012026-10-012025-05-01https://etd.aau.edu.et/handle/123456789/9053Deepfake technology has quickly grown through advances in deep learning and generative adversarial networks, enabling the creation of highly realistic synthetic media. However, these technological breakthroughs have also raised serious alarms concerning the possible misappropriation of DF content for spreading misinformation, fraud, and voice impersonation. Despite the growing complexity of Deepfake generation, there exists a considerable disparity in developing effective and robust detection techniques for audio Deepfakes, particularly for languages such as Afaan Oromoo. As widely spoken as Afaan Oromoo is despite the richness of linguistic content, insufficient focused research and customized detection models exist that are addressed to its characteristic phonetic and prosodic features. The identified gap reveals that speakers and systems utilizing Afaan Oromoo are vulnerable to risks linked to synthetic audio, such as security vulnerabilities and the dissemination of misinformation. This thesis proposes a novel hybrid deep learning approach for audio Deepfake detection. A hybrid model that integrates Convolutional Neural Networks (CNNs) for spatial feature extraction with Long Short-Term Memory (LSTM) networks for capturing temporal dependencies is developed. A comprehensive dataset was created by collecting genuine audio from reliable source and generating synthetic audio using a real-time voice cloning (RTVC) system. Extensive experiments demonstrated that the proposed CNN-LSTM model successfully learned the subtle acoustic and prosodic characteristics of Afaan Oromoo speech, achieving a training accuracy of 97.40% and an overall test accuracy of 95% a figure that accurately reflects the model’s robust generalization performance, as opposed to the previously noted 96% accuracy. The experimental results validated through additional metrics including precision, recall, and F1-score 0.95 for both real and fake. This research contributes a tailored, language-specific hybrid deep learning model for Deepfake detection, contribution a feasible solution to mitigate the risks associated with synthetic audio in underrepresented linguistic communitiesenDeepfakeAudio Deepfake DetectionHybrid CNN-LSTMDeep LearningAfaan OromooReal-TimeVoiceCloningThesis