Experiments on a Fast Mixture Density Likelihood Computation

Abstract

A computationally very expensive task arising within speech recognition systems using continuous mixture densities is the log-likelihood computation, which in our system essentially amounts to a nearest-neighbor search. A powerful nearest-neighbor search procedure is the Approximating and Eliminating Search Algorithm (AESA) [1], based on the triangle inequality. However, a direct application of this method is prohibitive in our framework, which is characterized by a high-dimensional feature space and a small set of prototypes per class. Thus, an improved triangle-inequality algorithm adapted to our conditions was derived. Experimental tests show that it is 35% faster than a straightforward search and 20% faster than the standard partial-distantce algorithm used in the Philips large vocabulary continuous-speech recognition system. The discussion particularly focuses on the influence of feature-space dimension and number of prototypes on the number of distances to be computed.