Influence of a Speaker’s Psycho-physiological State to His Voice Parameters and Results of Biometric Authentication by Speech Enabled Password

Abstract

In this work, two methods were used to calculate the identification characteristics of the speaker's voice. One of them is based on the direct Fourier transform, the second — on the window transformation with the subsequent integration of the values of each harmonic of all the windows. The information content of these characteristics is determined. An estimation is given of how the parameters of the voice and their informativeness change depending on the degree of alcoholic intoxication of a person and in a sleepy state. A computational experiment was carried out to evaluate the reliability of recognition of speakers in the space of selected features using functionals based on the Bayesian hypothesis formula, Pearson measure, chi-module measure, Gini criterion, Cramervon Mises, and perceptrons trained in GOST R 52633.5-2011, and networks of quadratic forms. An estimation is given of the stability of these methods and functionals to the psychophysiological state of the speaker in terms of the robustness of the obtained recognition results.