Improving discrimination of confusable words using the divergence measure

This paper describes the use of the divergence measure as a criterion for finding a transformation matrix which will map the original speech observations onto a subspace with more discriminative ability than the original. A gradient-based algorithm is also proposed to compute the transformation matrix efficiently. The subspace approach is used as a preprocessing step in a hidden Markov model (HMM) based system to enhance discrimination of acoustically similar pairs of words. This approach is compared with standard linear discriminant analysis (LDA) techniques and shown to yield as much as 4.5% improvement. The subspace approach is also applied successfully to a more general recognition problem, i.e., discrimination of K confusable words, using the average divergence measure.