Calculating likelihood ratios between speakers in subjective listening tests

Abstract

Voice identification using measurements of physical parameters, generally made through instrumental techniques, combines a significant performance, especially in the case of automatic methods, with an easy control of the data obtained. However, the possibility of working in this field with subjective investigation criteria on the basis of human listening should not be underestimated. In the present work we illustrate a method for obtaining a numerical value based on the likelihood, for subjective tests in speaker recognition tasks. This method was finally validated through a number of experiments made by two different types of listener: non-experts and technical experts.