Sixth European Conference on Speech Communication and Technology
A novel discriminative method for estimating the parameters of Hidden Markov Models (HMMs) is described. In this method, the parameter values are chosen to ensure that the characteristics of each sound class can be maximally separated. Compared with the significant method known as the Maximum Mutual Information (MMI) estimation, the novel method represented in this paper adopts a new kind of criteria called MSDI (Maximum Samples Distinction Information). It parries many computational problems in estimating iteration. The experimental results show that the hit rate can be raised by about 6 percent compared with the MLE (Maximum Likelihood Estimation), which is similar to the other experimental results based on MMI training.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Wu, Jian / Guo, Qing (1999): "A novel discriminative method for HMM in automatic speech recognition", In EUROSPEECH'99, 2761-2764.