EUROSPEECH '95

A computationally very expensive task arising within speech recognition systems using continuous mixture density HMMs is the loglikelihood computation. In the Philips largevocabulary continuousspeech recognition system it consumes 50%  75% of the decoding time. In our system the loglikelihood computation amounts to a nearestneighbor 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. In this paper, we show that a Hamming Distance Approximation (HDA) of the angles between the vectors leads to a powerful nearestneighbor search technique with negligible memory demands. Thus the likelihoodcomputation was sped up by a factor of 10 without significant increase in the word error rate of our large vocabulary speech recognizer. Since the likelihoodcomputation in this system consumed 66% of the recognition runtime, the overall decoding runtime could be reduced by a factor of 2.5. We also report results on Tldigits and the WSJ task.
Bibliographic reference. Beyerlein, Peter / Ullrich, Meinhard (1995): "Hamming distance approximation for a fast loglikelihood computation for mixture densities", In EUROSPEECH1995, 10831086.