ISCA Archive ICSLP 2002
ISCA Archive ICSLP 2002

A Gaussian selection method for multi-mixture HMM based continuous speech recognition

Raymond H. Lee, Eric H. C. Choi

This paper concerns improving Gaussian selection for reducing output probability computation. We investigate the use of principal component analysis (PCA) to generate questions for a decision tree which is then used to cluster a set of Gaussians for selection purpose. By dividing a feature vector into several subspaces and generating a decision tree for each subspace, we are able to generate a smaller shortlist and hence reduce computation further. Moreover we investigate different voting strategies to combine the shortlists selected from individual decision trees. Experiments on a Mandarin Chinese base syllable recognition task have revealed that our proposed method virtually does not degrade recognition accuracy, even though there is more than 50% reduction in output probability computation.