As deepfake audio becomes increasingly realistic, the need for reliable detection models is critical. This thesis evaluates the effectiveness of statistical versus deep learning methods for identifying deepfake audio in the for-norm release of the Fake-or-Real dataset. Statistical baselines are established using Logistic Regression and Random Forest models trained on handcrafted, non-learned acoustic features. The Random Forest model achieves a strong test accuracy of 0.815 and a test AUC of 0.888, showing that classical signal processing features remain competitive. Deep learning models are evaluated using log-Mel spectrogram inputs and convolutional architectures, including a lightweight CNN trained from scratch and a ResNet18 model fine-tuned for single-channel spectrograms. Threshold-based metrics are computed using a decision threshold selected on the validation set and then fixed for test evaluation on the held-out test set. Using this framework, the CNN attains a test accuracy of 0.745 with a test AUC of 0.882, while ResNet18 achieves the highest test AUC of 0.964 and a test AP of 0.963 but a lower test accuracy of 0.528 due to calibration mismatch between the validation and test splits. The results show that tree-based statistical baselines remain competitive in accuracy under fixed decision rules, while deeper architectures improve ranking performance and capture spectro-temporal cues that are not fully represented by handcrafted features, thereby highlighting the importance of calibration and threshold selection in deep learning models for reliable deployment.