Model Selection Criterion using Confusion Models for HMM Topology Optimization

Hidden Markov model (HMM) has been widely used in the area of speech and handwriting recognition, because of its excellent model power. If the number of parameters of HMM increases, the likelihood of in-class data tends to increase. At the same time, likelihood of out-of-class data also increases, so that excessive number of parameters diminishes discrimination power of HMM. In this paper, we proposed a new model selection criterion using confusion models, trained with confusion data in order to manage this problem. We built confusion models of the same number of parameters that standard models have. The proposed method, CMC (confusion model selection criterion), maximizes the modeling power of HMM while maintaining discrimination power as well, since the proposed method prefers standard models that output higher likelihood for the in-class data and confusion models that output lower likelihood for the out-of-class data. We performed handwriting recognition experiments using the CMC, and got better recognition accuracy using the propose method compared with ML and BIC