Markov Clustering Algorithm for Unsupervised Speaker Recognition System



The i-vector speech context representation became the state of the art in speaker verification. We had significant progress in this challenge with supervised methods (Cosine Distance with Linear Discriminante Analysis and Within Class Covariance method). However, the recent researches propose to use unlabeled data set of i-vectors to increase the size of training dataset and decrease the cost of collected data. This is why we base our study on the i-vectors space and work on unsupervised methods. In this study, we use a clustering method, the Markov Clustering process (MCL), to recognize natural groups of i-vectors which represent one speaker, within a class of entities. The MCL algorithm is a fast and scalable unsupervised cluster algorithm based on simulation of stochastic flow in graphs. The clustering result is used in the standard supervised speaker recognition system to evaluate the performances. We will compare these performances with other graph clustering methods, such as the Infomap, Self-Organizing Map and Newman-Girvan.