DATA REFINING HMM, A NEW APPROACH TO HMM BASED SPEECH RECOGNITION SYSTEM IMPROVEMENT

The accuracy of HMM based speech recognition systems is limited due to some HMM presumptions. In this paper, a decorrelation method named “data refining ” has been applied to HMM input sequence. The main idea in this study is to produce a nearly independent vector sequence containing all of the speech information in addition to consistency with model assumptions. In this paper, A few data refining methods were proposed and examined and the results were compared with each other and with standard HMM. Data refining method shows 3 % improvement on recognition rate and 30 % improvement on recognition computational cost, much better than other SSM methods