Locally Weighted Linear Discriminant Analysis for Robust Speaker Verification

Channel compensation is an integral part for any state-of-theart speaker recognition system. Typically, Linear Discriminant Analysis (LDA) is used to suppress directions containing channel information. LDA assumes a unimodal Gaussian distribution of the speaker samples to maximize the ratio of the between-speaker variance to within-speaker variance. However, when speaker samples have multi-modal non-Gaussian distributions due to channel or noise distortions, LDA fails to provide optimal performance. In this study, we propose Locally Weighted Linear Discriminant Analysis (LWLDA). LWLDA computes the within-speaker scatter in a pairwise manner and then scales it by an affinity matrix so as to preserve the withinclass local structure. This is in contrast to another recently proposed non-parametric discriminant analysis method called NDA. We show that LWLDA not only performs better than NDA but also is computationally much less expensive. Experiments are performed using the DARPA Robust Automatic Transcription of Speech (RATS) corpus. Results indicate that LWLDA consistently outperforms both LDA and NDA on all trial conditions.