Phonemes have characteristic properties stich as unique temporal structure, context sensitive behaviour and specific duration etc. Phoneme models should incorporate such constraints to provide better classification accuracy. In this paper these phonemic properties are incorporated into a HMM based phoneme recognizer with the addition of several degrees of freedom to the HMM state. The resulting models have shown improved performance on the TIMIT database.
Bibliographic reference. Sitaram, R. N. V. / Sreenivas, Thippur (1995): "On incorporating phonemic constraints in hidden Markov models for speech recognition", In EUROSPEECH-1995, 775-778.