Reduced universal background model for speech recognition and identification system

Abstract

Minimal Enclosing Ball (MEB) has a limitation for dealing with a large dataset in which computational load drastically increases as training data size becomes large. To handle this problem in huge dataset used for speaker recognition and identification system, we propose two algorithms using Fuzzy C-Mean clustering method. Our method uses divide-and-conquer strategy; trains each decomposed sub-problems to get support vectors and retrains with the support vectors to find a global data description of a whole target class. Our study is experimented on Universal Background Model (UBM) architectures in speech recognition and identification system to eliminate all noise features and reducing time training. For this, the training data, learned by Support Vector Machines (SVMs), is partitioned among several data sources. Computation of such SVMs can be efficiently achieved by finding a core-set for the image of the data. © 2012 Springer-Verlag Berlin Heidelberg.