The rapid advancement of deep learning technologies has led to the creation of highly convincing synthetic speech, raising significant concerns in areas such as digital security, media authenticity, and privacy. In response to this emerging threat, this work proposes a spectrogram-based machine learning framework for detecting deepfake audio. The approach involves converting audio signals into mel-spectrogram representations to capture essential time-frequency features, which are then used as input to three ensemble learning classifiers: MobileNetV2, DenseNet121, and EfficientNetB0. These models are trained and evaluated on the Deep Voice dataset, which comprises both real and synthetic voice samples. Experimental results demonstrate the effectiveness of the proposed method in distinguishing deepfake audio with high accuracy. This study underscores the potential of leveraging visual feature extraction techniques in combination with deep learning models for enhancing audio-based forgery detection systems.

Keywords: Deepfake Audio, Mel-Spectrogram, Audio Forensics, MobileNetV2, DenseNet121, EfficientNetB0, Deep Voice Dataset, Synthetic Speech Detection, Machine Learning, Ensemble Learning.