Improved speech modelling and recognition using a new training algorithm based on outlier-emphasis for nonstationary state HMM

In this study, we develop a modified maximum likelihood algorithm for optimally estimating the state-dependent polynomial parameters in the nonstationary-state HMM. The newly devised training method controls the influence of outliers in the training data on the constructed models. For an alphabet recognition task, outlier emphasis resulted in improved performance. An error rate reduction of 14% is achieved for the linear trend and 7.5% is obtained for the stationary-state HMMs over the conventional models trained by the Viterbi algorithm based on the joint-state maximum likelihood criterion. The properties of the nonstationary-state HMM trained with the proposed approach are analysed by examining goodness-of-fit of the real speech data to the polynomial trajectories in the model. (C) 1998 Elsevier Science B.V. All rights reserved.