In this paper, we propose an overall risk criterion for discriminative training of HMM-based speech recognizers using the minimum-error-rate classification of the Bayes decision theory. A reduced gradient method is applied to the proposed criterion for discriminative HMM parameter estimation. The resulting algorithm consists of gradient descent terms for competing classes and gradient ascent term for the correct class. In Korean digit recognition experiments, about 20 % reduction of the number of errors has been achieved.
Bibliographic reference. Na, Kyungmin / Jeon, Bumki / Chang, Dong-Il / Chae, Soo-Ik / Ann, Souguil (1995): "Discriminative training of hidden Markov models using overall risk criterion and reduced gradient method", In EUROSPEECH-1995, 97-100.