Robust speaker verification

Abstract

Smartphones are having high penetration rate in fast-growing countries. There are no doubt about how important they are in our daily lives. Together with the rising popularity of smartphones, come many security and privacy issues. Hence, to help solve this problem, biometric systems such as speaker recognition are introduced. In this project, a speaker recognition system was developed for Android platform, targeting a mainstream device – the Samsung Galaxy S3. A user-friendly application, namely “You Voice” was implemented, which allows users to train their own speaker models, and test them with any unknown voice. You Voice let the genuine speaker pass, while rejecting speeches from impostors. To ensure the accuracy and stability of the Android application, various experiments were conducted on PC. Applying Gaussian Mixture Model technology, a number of Universal Background Models were trained and tested. Experimental results showed that system performance achieved its peak at 256 GMM mixtures. In future, more experiments should be carried out, using better technologies such as GMM-SVM and i-Vector. The Android application could also be improved further for better user experience.