Investigation into the discriminating potential of cepstral coefficients for Forensic Voice Comparison (FVC)

Abstract

Forensic Voice Comparison (FVC) is the comparison of known suspect voice recordings with unknown offender voice recordings. The objective of this thesis is to determine the regions of real cepstrum coefficients (RCCs) that contain useful FVC information and to improve the FVC performance. In this research, RCCs are the input parameter set preferred over other cepstral coefficients (CCs), like linear prediction cepstral coefficients and Mel frequency cepstral coefficients for FVC. A new method called Principal Component Analysis Kernel Likelihood Ratio (PCAKLR) model developed at The University of Auckland is used for computation of likelihood-ratio (LR) values in this research. Researchers use lower order CCs for FVC as they believe higher order CCs do not contain much FVC information. In traditional methods for computing LR values like Univariate Analysis, Multivariate Kernel Density estimation and Gaussian Mixture Model – Universal Background Model do not allow researchers to use large numbers of input parameter. This thesis aims at investigating the possibility of speaker discriminating information being present in the regions of RCCs other than the lower order ones, applying PCAKLR model. Two strategies based on the pitch of speakers are identified to choose the regions of RCCs other than the lower order ones that might contain useful FVC information. Log-likelihood-ratio cost is used for computing the accuracy of FVC. values for these chosen regions are better than those using lower order RCCs in few experiments and suggests that these regions might carry some FVC information. In PCAKLR model, the first step is Principal Component Analysis (PCA) that transforms input parameter set into a new parameter set, the elements of which are ordered according to their information content. The strategies to remove the transformed parameters from PCA that contain little or no speaker discrimination information are identified. improves comparably when those transformed parameters with not much useful FVC information are discarded.