Dynamic Features for Segmental Speech Recognition

Abstract

Speech models and features that emphasise the dynamic aspects of speech can provide improved speech recognition. The cepstral time matrix has been established as a successful method of encoding dynamics. The paper extends this set of dynamic features, considering cepstral time features on both a segmental and subsegmental level. This offers the potential of using a conditional PDF for the state observation within a HMM and incorporating this into the training stage. Methods of linear discriminative analysis are applied to the new feature set to identify the subset of features making the greatest contribution to the task of recognition