A Hybrid MFCC–WavLM Feature Fusion for Audio Deepfake Detection

Abstract

Recent advances in generative artificial intelligence have enabled highly realistic speech synthesis using text-to-speech (TTS), voice conversion (VC), and neural voice cloning techniques, posing significant security threats to Automatic Speaker Verification (ASV) systems. Conventional handcrafted features such as Mel-Frequency Cepstral Coefficients (MFCCs) effectively capture spectral characteristics but may fail to detect sophisticated artifacts introduced by modern speech synthesis models. In contrast, self-supervised models such as WavLM provide rich contextual representations but may not fully capture complementary lowlevel acoustic information. This paper proposes a hybrid audio deepfake detection framework that combines an 80-dimensional MFCC feature vector with a 768-dimensional WavLM embedding to form an 848-dimensional hybrid representation. The extracted features are standardized using StandardScaler, balanced using the Synthetic Minority Over-sampling Technique (SMOTE), and classified using a Deep Neural Network (DNN). Experiments on the ASVspoof2019 Logical Access (LA) dataset show that the proposed framework achieves 94.8% classification accuracy and a weighted F1-score of 0.95 on a 5,000-sample evaluation subset, demonstrating the effectiveness of the hybrid representation for spoofed speech detection.