Application of an Annular/Sphere Search Algorithm for Speaker Recognition

Abstract

In this work, an alternative search algorithm for vector quantization codebook is applied as a way to improve the performance of an automatic speaker recognition system. The search algorithm is based on geometrical properties of the vector space, defining annular and spherical regions instead of a full search method. The speaker recognition system is intended to identify a suspect, between a small group of persons, using low quality recordings, working as a text independent automatic speaker recognition system. Because the rate of recognition required in forensic applications is extremely important, the use of good discrimination algorithms can reduce the risk of bad decisions. The performance of the system under such a conditions is reported. Besides the few speaker samples used for training, a high recognition rate was obtained, so it was found an improvement of the recognition rate over the full search method.