Incorporating phonetic properties in hidden Markov models for speech recognition

In this paper the incorporation of important phonetic properties into hidden Markov models (HMM) is studied. Phones have characteristic properties such as unique temporal structure, context sensitive behavior and specific duration, etc. New HMMs which incorporate the above properties with additional degrees of freedom to the standard HMM states are proposed. The use of each of the phonetic property for speech recognition is demonstrated using the new HMMs. All the algorithms required for using these new models in various applications of speech recognition have been presented. Experimental comparison with the standard discrete HMM for a speaker-independent continuous speech phone recognition task show that consistent improvement is achieved by the new models