A reduced search algorithm for speaker recognition

Abstract

In this work, a reduced search algorithm for vector quantization codebooks is applied as a way to reduce the risk of wrong decisions in an automatic speaker recognition system. Instead of a full search method, the algorithm is based on the geometrical properties of the vector space, reducing the search to those codebooks which are closer to the vector under test. 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. It was found that the alternative search algorithm can be used to reduce the risk of wrong decisions, which is specially important in forensic applications.