Automatic selection of phonemes from an equally spaced quasi-phoneme string by the entropy principle

A new method to perform the phonemic segmentation of speech is introduced. The proposed method is based on analyzing the equally spaced quasi-phoneme string produced by the acoustic preprocessing and the phonemic labeling. The string consists of the most probable phonemic candidates and their reliabilities. The probabilities of different phonemes and the entropy of the string are evaluated at each analysis interval. The minima of the entropy indicate the steady-state regions of the phonemes. The performance of the entropy method is compared with that of two other segmentation methods.