Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Fast Log-Likelihood Computation for Mixture Densities in a High-Dimensional Feature Space

Peter Beyerlein

Philips GmbH Forschungslaboratorium Aachen, Aachen, Germany

A computationally very expensive task arising within speech recognition systems using continuous mixture density HMMs is the log-likelihood computation. In the Philips large vocabulary continuous-speech recognition system it consumes 50%-75% of the computing resources. In our system the log-likelihood computation amounts to a nearest-neighbor search, i.e. to a search for the component density of a mixture density whose mean vector has a minimal distance to the observed feature vector. Fast nearest-neighbor search techniques based on the triangle inequality are very powerful if the dimension of the feature space is lower than about 10. However, a direct application of these techniques is prohibitive in our framework which is characterized by a high-dimensional feature space and a small number of component densities per mixture density. In a typical setup we have 120 component densities per mixture density and a dimension of 63. This paper

Full Paper

Bibliographic reference.  Beyerlein, Peter (1994): "Fast log-likelihood computation for mixture densities in a high-dimensional feature space", In ICSLP-1994, 271-274.