Feature selection methods for hidden Markov model-based speech recognition

In the paper three different feature selection methods applicable to speech recognition are presented and discussed. Widely known approaches, like the principal component analysis, discriminant feature analysis and sequential search methods, have been customised for the use with a hidden Markov model based classifier. When comparing the methods we focus mainly on their ability to reduce the size of the feature vectors standardly used in speech processing. It is demonstrated that the sequential methods and the discriminative analysis are well suited for that task. Both of them may contribute to a recognition time reduction by a factor higher than two without a significant loss of accuracy, particularly, in the combination with a two-level classification scheme.