Spectrogram Features and Xception Classifier Based Audio Spoof Detection System
Nidhi Chakravarty, Mohit Dua · 2024
Voice biometric authentication has gained significant attention in recent years due to its non-intrusive and user-friendly nature. In this research paper, we present a comprehensive study on the effectiveness of three different spectrograms - Constant Q Transform (CQT), Mel and Gammatone at the front end of an Audio Spoof Detection (ASD) system. The back end of the system employs the Xception pretrained model. Our experimental results reveal that Gammatone Spectrograms outperform both CQT Spectrograms and Mel Spectrograms in terms of performance. Specifically, combination of Gammatone Spectrogram with Xception pretrained has achieved a remarkable 3.5% and 2% Equal Error Rate (EER) for ASVspoof 2021 DF and DECRO dataset, respectively