Decision Procedures for Discriminating between Speakers

Abstract

Decision procedures are considered in which one of a number of alternative points must be identified as belonging to the same set as a given stimulus point. If the points are all described by values on n characteristics, the decision may be made by calculating the distance from each of the alternatives to the stimulus in the n-dimensional feature space. If the features are not independent or not equally useful, the decision procedure may be improved by applying a linear transformation to the space before calculating distances. Since the decisions are based upon single-point representatives of sets rather than sets of points, standard methods of obtaining optimal transformations are not useful. Possible criteria of optimality, useful classes of transformations, and methods of obtaining good transformations are discussed. Transformations are applied to data in which expert listeners rated sentences from different speakers on 6 voice characteristics. The inability of diagonal transformations to improve this decision procedure is related to the nature of the speaker discrimination task. [Work supported by Decision Sciences Laboratory, ESD, U. S. Air Force Systems Command, Bedford, Massachusetts.]