In this paper, we present online text-independent speaker verification system developed at IIT Madras. Unlike, parametric approaches to model the distribution of speaker-specific features, we explore nonparametric approaches to capture the distribution of speaker-specific features. The distribution capturing ability of NonLinear Principal Component Analysis (NLPCA) neural network is exploited to build speaker models. In this study, linear prediction cepstral coecients are used to represent speaker-specific information. Background normalization is performed using an rank based approach. The effectiveness of this speaker verification system is demonstrated on a database of 15 speakers.