Improved unsupervised NAP training dataset design for speaker recognition

The Nuisance Attribute Project (NAP) with labeled data provides an effective approach for improving the speaker recognition performance in the state-of-art speaker recognition system by removing unwanted channel and handset variation. However, the requirement for the labeled NAP training data may limit its practical application. In our previous study, a simple unsupervised clustering algorithm based on dot products between supervectors was introduced for designing NAP training dataset without a prior knowledge about channel and speaker information. Using such clustering results as the initial training dataset, in this paper, we make a further improvement of the training dataset by enhancing similarity measurement of supervectors via NAP projection and score normalization. The effectiveness of this unsupervised NAP training dataset design strategy has been verified in the experiments using the in-house development dataset of IIR submission for the 2012 NIST SRE.