A CONTEXT ADAPTATION APPROACH FOR BUILDING CONTEXT DEPENDENT MODELS IN LVCSR

摘要

This paper introduces a new context adaptation framework for building context dependent HMM models in LVCSR. In this new framework, all states of each center phone are clustered into groups by the decision tree algorithm. All the tied states of context dependent HMM models were then derived by adapting the parameters of the multiple-mixture context independent model via data dependent MAP (maximum a posteriori probability)method using the training vectors corresponding to the tied state. An advantage of this approach is that it can maintain a high prediction and classification power given limited training data therefore the model trained in this framework is more reliable than in conventional framework. Experimental results on Wall Street Journal corpora demonstrate that the proposed approach leads to a significant improvement in recognition performance.