Improving hidden Markov model performance in phoneme classification by fuzzy smoothing
Farbod Hosseyndoost, Mohammad Teshnehlab · 2005
In this paper, two kinds of uncertainties, in speech production and recognition, are introduced. It is shown that one class of these uncertainties can be best understood by the notion of probability while the other could be described as fuzziness. Based on the given concepts, a new method of fuzzy smoothing is proposed. The goal of this method is to make transition from one state to another gradual, so that contribution of each observation to the emission probability becomes fuzzy. In addition, an innovative implementation method is suggested. It is shown that adding a length normalized time dimension to the feature space can serve as fuzzy smoothing. Error rates in phoneme classification are compared on TIMIT database. Results show, a significant improvement in classification rate over the original HMM, while imposing no computational intricacy.