Enhancing Spoof Detection in Automatic Speaker Verification Using CQCC Optimization and ViT Architecture

Spoof detection is found to be essential for improving the security features of automatic speaker verification (ASV) systems, which are primarily used in authentication. The primary goal of this study is to enhance the performance and efficiency of spoof detection using speech samples taken from the ASVspoof 2019 dataset. The Constant Q Cepstral Coefficients (CQCC) extracted from these speech samples act as an important key feature. Feature optimization methods such as Genetic Algorithm (GA), Grey Wolf Optimizer (GWO), and Mayfly Optimizer (MO) are used to refine these features and hence enhance the model accuracy with minimal time cost. A Vision Transformer (ViT) model is then trained using each optimized feature, and the performance is evaluated by comparing the results from different optimization methods. Time analysis shows a substantial reduction in training time per epoch when the optimized features are used. The Genetic Algorithm attained the best performance, with a test accuracy of 97% and the least training time. Equal Error Rate (EER) and the Tandem Detection Cost Function (t-DCF) are used as the evaluation metrics. This study demonstrates how feature optimization helps to enhance spoof detection accuracy while reducing processing time, hence becoming an authentic solution for real-time ASV systems.