Speaker Recognition Method Based on Phone N-gram Pruning and KPCA

Abstract

In order to solve the problem of disturbance due to data sparsity for the baseline phone n-gram system, a method based on Phone N-gram pruning and KPCA is brought forward. The phone n-gram with low probability is firstly pruned in the phone n-gram super vector. The kernel principal component analysis (KPCA) is then adopted to remove the bias which is brought about due to data sparse. When applying this method to the NIST 2006 speaker recognition evaluation (SRE) database, experimental results shows that a relative reduction of up to 29% in Error Equal Ratio (EER) is achieved over the previous baseline phone n-gram system.