Joint quantizer design and parameter estimation for discrete hidden Markov models

An approach that involves designing a vector quantizer to maximize the mutual information between the hidden Markov model (HMM) states and the quantized observations is presented. The iterative design of the quantizer and the HMM parameters is shown to be jointly a maximum-likelihood estimate. Methodologies for using the maximum mutual information (MMI) criterion for quantizer design are described, and some initial results are presented to demonstrate that the MMI criterion yields improved speech recognition performance.<
>