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 introduces an efficient nearest-neighbor search algorithm adapted to the conditions of a high dimensional feature space and sparse data, gives a theoretical explanation and an experimental validation for the constraints of fast nearest-neighbor search techniques in a high-dimensional space.
Cite as: Beyerlein, P. (1994) Fast log-likelihood computation for mixture densities in a high-dimensional feature space. Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994), 271-274, doi: 10.21437/ICSLP.1994-71
@inproceedings{beyerlein94_icslp,
author={Peter Beyerlein},
title={{Fast log-likelihood computation for mixture densities in a high-dimensional feature space}},
year=1994,
booktitle={Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994)},
pages={271--274},
doi={10.21437/ICSLP.1994-71}
}