Hybrid speech recognition system with discriminative training applied for Romanian language
Inge Gavăt, M. Zirra, Oana G. Cula · 2002
This paper describes a hybrid connectionist-statistical system consisting of a neural network integrated in a hidden Markov model (HMM). The neural network used is the multilayer perceptron (MLP) and that network is the mechanism that computes the a-posteriori probability of a sequence of HMMs states. The classifier is based on total scores computed by Viterbi alignment for each hybrid model corresponding to the words in the vocabulary. Because of the lack of discrimination between the models and the unintended discrimination between the states in each model, we propose a solution that improves the system, namely an additional training task based on a cost function that approximates the misclassification rate of the hybrid system. The optimization criterion is based on a descent algorithm and the result is a minimum classification error. Our experiments on a 35 word vocabulary, show an improvement of the recognition rate from 92.4% for the case of a statistical system based only on HMMs, to 94.7% for the case of a hybrid HMM-MLP system, and to 97.9% for the case of an improved hybrid system with an extra discriminative training.