Towards making better hybrid pattern classification design for speaker identification

Abstract

Recent development on classify speaker data from a group of speaker is still insufficient to provide a satisfied result in achieving high performance pattern classification engine. There are two main difficulties in this field: how to maintain accuracy rate under incremental amounts of training data and how to reduce the time processing in the case embedded systems need to consider about efficient and simplicity of calculation. Recently we have proposed three difference hybrid pattern classification approach for text independent speaker identification system; in these approaches, we combined a hybrid GMM/VQ and decision Tree model. In this paper, we extend our investigations in order to select the most suitable hybrid GMM/VQ+DT model for real time application. For the first proposed hybrid modeling, both VQ model and GMM model will run parallel after signal preprocessing process; while the second type of proposed hybrid modeling, we present the use of decision tree in VQ techniques. The third method are extended from the second hybrid modeling which is using VQ decision rules for Gaussian mixture modeling in order to simplified the process. Experimental result shows that the third type of hybrid modeling should be considered for real world application due to the superior performance of time processing.