Fast fuzzy clustering approach for real time speaker identification with feature projection and dimension reduction

Abstract

Speaker identification applications are the highly commercialized during the speaker recognition and voice biometrics applications which experienced high interests and an increasing market during the last few years. Like other applications it has been experiencing an increasing market and investor interest [1]. However, it is still an open area for research. In this thesis we considered creating a real-time speaker identification system as an optimization problem. However, many researches proceeded in this direction. In our proposed system, pattern recognition techniques like clustering and p-tree search were used to reduce the number of calculations and database comparisons in order to speed up the identification process while entropy and principal component analysis were applied to reduce the dimensionality of data and thus reducing data storage size. These modifications allow the user making trade off between accuracy, response time and resource usage based on the usage case of action. Here, we took a newly proposed technique proposed by Karpov [33] to compare our technique with it. The results, when testing both systems on ELSDSR voices database, show that the newly proposed technique is better than the old proposed one in terms of timing, memory usage and accuracy.