Novel application of k-means support vector data description to speaker verification without background models

In this paper, we propose a speaker verification system without background models using a kmeans support vector data description. Support vector data description (SVDD) is an outlier detection method that is suitable for speaker verification since speaker verification involves detecting the input voice that is not from the desired speaker and rejecting an unauthorized speaker. However, in practice, it is not easy to use the SVDD for speaker verification since the time complexity increases rapidly as the amount of input data increases. Therefore, we propose the use of the k-means SVDD, which can reduce the volume of computation required, by using a divide-and-conquer strategy. In addition, a k-means algorithm can give a more detailed description of the input speech patterns because speech pattern classes consist of many sub-classes, which are similar to phones. Thus, we can achieve a better result in a speaker verification task by using the YOHO database.