An efficient scoring algorithm for Gaussian mixture model basedspeaker identification

摘要

This article presents a novel algorithm for reducing thencomputational complexity of identifying a speaker within a Gaussiannmixture speaker model framework. For applications in which the entirenobservation sequence is known, we illustrate that rapid pruning ofnunlikely speaker model candidates can be achieved by reordering thentime-sequence of observation vectors used to update the accumulatednprobability of each speaker model. The overall approach is integratedninto a beam-search strategy and shown to reduce the time to identify anspeaker by a factor of 140 over the standard full-search method, and byna factor of six over the standard beam-search method when identifyingnspeakers from the 138 speaker YOHO corpus