A Minimum Discrimination Information Approach for Hidden Markov Modeling

Commonly used approaches for hidden Markov modeling of information sources assume that the measurements were generated by some hidden Markov source, and attempt to find a maximum likelihood (ML) estimate of the parameters of that source. This assumption is, however, not necessarily true, especially for speech signals for which such models have been recently extensively applied. We propose an alternative approach for doing the modeling in which the model and the measurements are matched in an information theoretic way.