Integrated approach to speaker recognition in forensic applications

Abstract

Speaker recognition for forensic use should exhibit a high level of reliability. Considering the large differences in transmission and recording conditions for rest and reference speech samples, no single method of speaker recognition can provide the required level of decision confidence. To solve this problem, the authors propose a multistage approach consisting of a parallel application of aural-perceptual, visual and automatic recognition procedures in open sets of speakers. The first method is based on perceptual comparison of speech segments of an unknown speaker (i.e., the criminal) with the same or similar segments of known speakers (i.e., the suspects). The second method is based on visual comparison of sets of patterns corresponding to the same speech segments produced by the same speakers. These patterns are based on sonagrams and other specially extracted parameters, such as spectra in stationary segments of speech, distributions of time intervals between zero crossings, etc. The automatic methods are based on a probabilistic model of speakers' voices in open sets and nonparametric models with an application of dynamic time warping. This integrated approach was rested in practical forensic applications, e.g. in the identification of extortionists. In the case of very different technical conditions, only the integrated approach permitted positive identification decisions to be reached. Illustrations of the results for a real life case also are included.