IMPROVED DESIGN OF DTW AND GMM CASCADED ARABIC SPEAKER

Abstract

In this paper, we discuss about the design, implementation and assessment of a two-stage Arabic speaker recognition system, which aims to recognize a target Arabic speaker among several people. The first stage uses improved DTW (Dynamic Time Warping) algorithm and the second stage uses SA-KM-based GMM (Gaussian Mixture Model). MFCC (Mel Frequency Cepstral Coefficients) and its differences form, as acoustic feature, are extracted from the sample speeches. DTW provides three most possible speakers and then the recognition results are conveyed to GMM training processes. A specified similarity assessment algorithm, KL distance, is applied to find the best match with the target speaker. Experimental results show that text-independent recognition rate of the cascaded system reaches 90 percent.