Audio Deepfakes: Feature Extraction and Model Evaluation for Detection

Cutting-edge AI-driven tools are currently employed for replicating human voices, leading to the emergence of audio deepfakes. Initially designed to enhance experiences like audio-books, the potential misuse of audio deepfakes poses significant risks to public safety. Despite advancements in detecting video deepfakes, identifying audio deepfakes remains challenging. We employ methods to extract features such as Mel-frequency cepstral coefficients, chromagrams, spectral contrast, tonnetz, and Mel-spectrograms from audio samples. The extracted deep features are concatenated together and then employed in machine learning and deep learning models to evaluate the ASVspoof2021 dataset to categorize it into bonafide and spoofed speech utterances. The experimental results reveal that the multilayer perceptron model achieves the highest accuracy rate of 96.073%, while SVM, k-NN, XGBoost, and RF produce accuracy rates of 93.58%, 94.58%, 93.544%, and 93.58%, respectively. Additionally, DL models, including CNN, achieve an accuracy rate of 98.2%, DNN of 97.6%, and LSTM of 97.7%, showcasing the superiority of CNN in identifying deepfake audio amidst the evolving landscape of synthetic media.