Speaker Dependent Word Recognition based on Dempster–Shafer Theory using Linear Predictive Coding

In this paper we propose a new approach to short-time speaker dependent word recognition based on Dempster-Shafer theory using Linear Predictive Coding (LPC) coefficients. For this we used a database of ten pre-determined signals and one-incoming signal, these signals are generated from ten different words of ten-different persons by using LPC techniques. Now by measuring similarity between incoming signal and predetermined signals from the database, the recognition of a particular word is done. Correlation and monogenic signatures are two theories used to study the similarity of the two signals using discrimination factor. However mutual information theory predicts the probability of occurrence of a signal by measuring the information obtained in the signal. The values from these three theories are changed into mass functions and these are used as evidences in the Dempster-Shafer theory of evidence. All these evidences are combined using the Dempster’s rule and finally the best matching speaker dependent word is identified. Key-words: Linear Predictive Coding, correlation, monogenic signatures, mutual information, DempsterShafer theory of evidence.