Deepfake audio, which artificially replicates real voices, poses serious risks such as fraud, misinformation, and identity theft. While most research focuses on detecting deepfake videos, audio detection remains a relatively unexplored area. This study uses Mel-Frequency Cepstral Coefficients (MFCC) feature extraction, which mimics human auditory perception, to differentiate between real and synthetic speech. The Fake-or-Real dataset was used to train and evaluate multiple machine learning models. Among the tested models, the Support Vector Machine (SVM) performed best for short audio clips, while Gradient Boosting delivered better results for longer recordings. Additionally, the VGG-16 deep learning model achieved the highest accuracy of 93% on complex datasets. These results suggest that MFCC-based feature extraction combined with machine learning models can effectively detect deepfake audio.