I-Vector Selection for Effective PLDA Modeling in Speaker Recognition

Abstract

Data selection is an important issue in speaker recognition. In previous studies, the data selection for universal background model (UBM) training and for the background dataset of support vector machines (SVM) have been addressed. In this paper, we address the data selection for a probabilistic linear discriminant analysis (PLDA) model which is one of the state-of-the-art methods for i-vector scoring. We first show that the data selection using the conventional k-NN method indeed improves the speaker verification performance. We then propose a robust way of selecting k by using a local distance-based outlier factor (LDOF). We name our method as flexible k-NN or fk-NN. Our fk-NN obtained significant performance improvements on both male and female trials of the NIST speaker recognition evaluation (SRE) 2006 core task, NIST SRE 2008 core task (condition-6) and NIST SRE 2010 coreext-coreext task (condition-5).