Adaptive Ensemble Classification for Speaker Verification

Authors

Faculty of Electrical and Computer Engineering, University of Birjand, Khorasan Jonoobi, Iran

Abstract

This paper considers the ensemble classification for the text independent speaker verification issue. Using one classifier for the speaker verification may not result in dependable decision, because it may not exploit different characteristics of speech signal. Therefore, state-of-the-art speaker verification systems use an ensemble of classifiers for the verification. Most of the ensemble speaker verification systems use a weighted summation of the score of the individual expert classifiers to calculate the final score of the verification. The weights of this score fusion is obtained using a method, e.g. logistic regression, in the training phase. These works do not efficiently take into account issues such as correlation of classifiers and instance specific behavior of the base classifiers into account. In this paper a new solution is proposed for these two issues by using the process of ensemble design and combination rule based on training data. The obtained results on NIST 2004 speaker evaluation corpus show the effectiveness of the proposed methods in comparison to the sparse classifier fusion, as a baseline method.

Keywords

Files

Share

How to cite

Statistics