Gaussian selection algorithm in Continuous Speech Recognition

Abstract

Clustering of Gaussian mixture components, i.e. Hierarchical Gaussian mixture model clustering (HGMMC) is a key component of Gaussian selection (GS) algorithm, used in order to increase the speed of a Continuous Speech Recognition (CSR) system, without any significant degradation of its recognition accuracy. In this paper a novel Split-and-Merge (S&M) HGMMC algorithm is applied to GS, in order to achieve a better trade-off between speed and accuracy in a CSR task. The algorithm is further improved by introducing model selection in order to obtain the best possible trade-off between recognition accuracy and computational load in a GS task applied within an actual recognition system. At the end of the paper we discuss additional improvements towards finding the optimal setting for the Gaussian selection scheme.