Nonparametrically Trained Probabilistic Linear Discriminant Analysis for i-Vector Speaker Verification

In this paper we propose to estimate the parameters of the probabilistic linear discriminant analysis (PLDA) in text-independent i-vector speaker verification framework using a nonparametric form rather than maximum likelihood estimation (MLE) obtained by an EM algorithm. In this approach the between-speaker covariance matrix that represents global information about the speaker variability is replaced with a local estimation computed on a nearest neighbor basis for each target speaker. The nonparametric between- and within-speaker scatter matrices can better exploit the discriminant information in training data and is more adapted to sample distribution especially when it does not satisfy Gaussian assumption as in i-vectors without length-normalization. We evaluated this approach on the recent NIST 2016 speaker recognition evaluation (SRE) as well as NIST 2010 core condition and found significant performance improvement compared with a generatively trained PLDA model.