HybridSpec-MFCC: A Multi-Modal Approach for Robust Deepfake Audio Detection Using Spectral-Temporal Feature Fusion
Abstract
The use of AI-generated synthetic audio presents some new security issues and creates validity issues with respect to the content. This paper describes a deepfake audio detection system called HybridSpec-MFCC that depends on MFCCs, spectral contrast and chromagram features in a combined convolutional-recurrent architecture, as well as an attention module that is self-calibrating. Evaluation of the model was based upon the ASVspoof 2019 dataset, where it performed with 92.4% precision, 91.6% recall and 92.0% F1-score, it outperformed other existing methods and retained robust performance over different acoustic conditions. It was concluded that multi-modal fusion of spectro-temporal features yields a better and more generalizable performance for deepfake audio detection than a single feature.