Safeguarding Audio Integrity: Ensemble Architectures for Real-World Fake Audio Identification
Soundarya Lahari Kasturi, S. Nivetha, B. Selva Rani, Subbiah Vairamuthu, Sourabh Tiwari, Devendra Yadav, Rashmi T Shankarappa · 2024
Through this research, we introduce a unique Audio Spoof Detection System for distinguishing between actual, fraudulent, and generated audio recordings. Strong detection algorithms are required since the spread of audio modification techniques poses a major threat to the integrity of audio-based systems. Combination of 7 key voice features from cepstral coefficients and spectrograms were used to cover all possible spoofing. We are proposing ensemble methods to enhance the accuracy of conventional machine learning models such as Random Forest, State Vector Classifier (SVC), K-Nearest Neighbors (KNN) etc. and deep learning models such as Recurrent Neural Networks (RNN), Time-Delay Neural Networks (TDNN), SqueezeNet etc. The proposed model is evaluated using the ASVspoof 2021 logical access (LA) and Deep Voice dataset. Experimental results show that our proposed model significantly elevates performance compared to the baseline and state-of-the-art models.