A novel loss function for the overall risk criterion based discriminative training of HMM models

In this paper, 1 we propose a novel loss function for the overall risk criterion estimation of hidden Markov models. For continuous speech recognition, the overall risk criterion estimation with the proposed loss function aims to directly maximise word recognition accuracy on the training database. We propose reestimation equations for the HMM parameters, which are derived using the Extended Baum-Welch algorithm. Using HMM, trained with the proposed method, a decrease of word recognition error rate of up to 17.3% has been achieved for the phoneme recognition task on the TIMIT database.