Neural networks based channel compensation for i-vector speaker verification

Abstract

Linear discriminant analysis (LDA) and Gaussian probabilistic LDA (PLDA) have been shown to effectively suppress channel- and session-variability of i-vectors. But they suffer the following limitations: 1) In LDA, a single linear transformation may not be adequate to describe the nonlinear relationship of features and 2) Gaussian-PLDA assumes the speaker and channel factors follow a Gaussian distribution, but they are actually non-Gaussians. We consider neural networks (NN) as a way to overcome the limitations, that captures the nonlinear relationship of features and does not require prior assumptions. This paper investigates three NN based channel compensation methods: deep metric learning, NN classifier, and deep denoising autoencoder and compares their performance with LDA and PLDA. Experiments conducted on NIST 2010 speaker recognition evaluation suggest that NN-based channel compensation methods are superior to LDA and that the performance of NN classifier is better than that of PLDA under most of common evaluation conditions. Additionally, this paper also helps us understand the relationships among LDA, PLDA, and NN based methods.