A fuzzy-GMM classifier for multilingual speaker identification

Abstract

In this paper, a new modeling approach is proposed by hybriding the features of expectation-maximization algorithm(GMM) and fuzzy c-means algorithm(FCM). Based on the analysis over conventional GMM technique, we suggested a new speaker identification system by fusing GMM (optimized using EM algorithm) and FCM, to improve the identification rate further in multilingual speaker identification system. The proposed technique and GMM technique was evaluated in mono and multilingual environments. Experiments were done also by varying the initial code books for generating speaker model. The experimental result shows improvements on a combined FGMM system, which employs fusion for the multilingual context with varying initial code books gives an improvement of minimum 2.98% than existing GMM approach. MFCC technique is used for extracting the features. The algorithms were compared using TIMIT database of 54 speakers speaking 3 languages like English, Hindi and Tamil.