Secure Speaker Verification at Web Login Page Using Cepstral Features

摘要

The recent development of online activity increases the need of security in the web login page. The login page using speaker verification can be use as high secure user verification method for web application. In this research, the speaker recognition system on web page has successfully built for login authentication security. This system consists of digital signal processing to extract the speaker features in Mel Frequency Cepstral Coefficient (MFCC), Nonlinear Power Spectral Subtraction to filter the speech signal from contaminating noise, Gaussian Mixture Model (GMM) to imitate the voice tract model, K-Means and Expectation-Maximation (EM) algorithm for training the model. In order to improve the security level, the system uses Secure Socket Layer (SSL) with 1024 bits RSA encryption. From this research, we have succeeded in optimizing the signal quality up to 5 dB SNR, the mean error recognition level of FAR about 23.3% and FRR 27.5 and the maximum accuracy of recognition system around 88% when the quality of speech signal is clean. The computation time for enrolment is about 552573.5 milliseconds and for verification about 129062.6 milliseconds.