Accurate, Fast and Low Computation Cost of Voice Biometrics Performance using Model of CNN Depthwise Separable Convolution and Method of Hybrid DWT-MFCC for Security System

Identity theft, a pervasive criminal risk in the digital realm, particularly in online transactions, demands innovative security solutions. Voice biometrics, a cutting-edge technology, have been developed to ensure the protection of one's identification. This study, a significant step forward, focuses on the development of voice biometrics using deep learning, specifically CNN Depthwise Separable Convolution (DSC) and CNN Residual. The research on these two systems was conducted to determine accuracy, performance evaluation, computing load, and training process time for effectively, rapidly, and accurately verifying user voice for banking transaction security. The initial CNN residual test yielded a high validation accuracy of 98.6345%. However, the large number of CNN residual parameters resulted in a training time of 7.37 seconds, increasing the computational workload. The second CNN DSC test exhibited a high validation accuracy of 98.3542%. The CNN DSC was successful in decreasing the parameter count, resulting in a reduction of 5.12 seconds in training time. Upon analyzing the test results, it is clear that the CNN DSC has superior performance, resulting in faster training times and less memory consumption. This effectively addresses the problem of high computational costs and significantly enhances user identity security in banking transactions, a crucial aspect of modern banking.