On Leveraging Conversational Data for Building a Text Dependent Speaker Verification System

Recently we have investigated the use of state-of-the-art text independent and text-dependent speaker verification algorithms for a text-dependent user authentication task and obtained satisfactory results mainly by using a fair amount of text-dependent development data. In our study, best results were obtained using the NAP framework rather than using the more advanced JFA and i-vector-based frameworks. In this work we investigate the ability to build high accuracy i-vector-based systems by leveraging widely available conversational data. We explore various techniques for transforming conversational sessions in such a way that attributes which are more relevant to the text-dependent task are enhanced. Using these techniques we managed to reduce verification error significantly.