Discriminant Analysis Methods Comparison in I-Vector Space for Speaker Verification

Identity vectors are the state-of-the-art feature vectors for speaker recognition applications. One of the most important advantages of i-vector is its allowance for implementation of channel and noise compensatory methods such as linear discriminant analysis (LDA). The motivation for this is to look for new orthogonal axes to achieve superior discrimination between different classes. The axes should comply with the inter-class variance maximization and intra-class variance minimization requirements. The conventional method for the LDA transform computation considers Gaussian distribution assumption and uses parametric representations for both intra-and inter-speaker scatter matrices. Of course, the actual distribution of i-vectors may not necessarily be Gaussian. In this paper, we investigate the performance of LDA, and three nonparametric techniques, i.e., NDA, GDA, and SVDA separately and in combination with LDA. Experiments were conducted on TIMIT and NIST SRE 2008 datasets with MFCC and PNCC feature vectors. The results show that using the combination of parametric and nonparametric methods can lead to better results.