Development and evaluation of discriminative training algorithm using HMMs with mixtures of linear trended functions

We extend the maximum likelihood (ML) training algorithm to the minimum classification error (MCE) training algorithm for discriminatively estimating the state-dependent and mixture-dependent polynomial coefficients in the linear trended HMM. The main motivation of this extension is a greater degree of freedom in the modeled trajectory space associated with the trended HMM than that with the conventional, stationary-state HMM which describes only the (piecewise) constant “degraded trajectories” in the observation data. Hence, the discriminative training implemented for the trended HMM should have a strong power to disambiguate the observational trajectories generated from nonstationary sources corresponding to different speech classes. The properties of the MCE formulation for training the trended HMM is analyzed by examining the goodness-of-fit of the raw speech data to the polynomial trajectories in the model. Phonetic classification results are reported which demonstrate consistent performance improvements with use of the MCE-trained trended HMM both over the regular ML-trained trended HMM and over the MCE-trained stationary-state HMM