Image similarity functions in non-parametric algorithms of voice identification

Abstract

This paper is dedicated to the question of the choice of a function of similarity between images in non-parametric alogorithms of voice recognition. The usefulness of 10 similarity functions (8 distances and 2 nearness'es) in three non-parametric identification algorithms – NN (nearest neighbour), k-NN (k-nearest neighbours) and NM (nearest mean) – was investigated for three sets of parameters (1 natural and 2 normalized). Results obtained for a population of speakers from a closed set with size M = 20 (after 10 repetitions of the learning and test sequences) have proved that the Camberr distance function prevails in all types of parameters and algorithms. Other functions ensure a differentiated discrimination force strongly dependent on the algorithm and form of parameters. Limited usefulness of the square of Mahalonobis distance in comparison to other similarity functions was proved, as well as generally worse results for the NM algorithm.