Text independent speaker identification with finite multivariate generalised Gaussian mixture model and k-means algorithm

Abstract

In this paper, we propose text independent speaker identification with a finite multivariate generalised Gaussian Mixture Model (GMM) with a k-means algorithm. Each speaker's speech spectra are characterised with a mixture of generalised Gaussian distribution that includes Gaussian and Laplacian distribution as a particular case. Speech analysis is done with the Mel Frequency Cepstral Coefficients (MFCC) extracted from the front end process. Using the EM algorithm and k-means algorithm the model parameters the numbers of acoustic classes associated with each speech spectra are Determined . The performance of the proposed algorithm is studied through experimental evaluation and observed that this algorithm outperforms the existing speaker identification algorithm with GMM. It is also observed that this algorithm performs efficiently even with a heterogeneous population with small (less than 2 seconds) utterances.Keywords: gaussian mixture model; GMM; generalised gaussian mixture model; MFCC; mel frequency cepstral coefficients; k-means clustering; EM algorithm and speech analysis.Reference to this paper should be made as follows: Sailaja, V., Srinivasa Rao, K. and Reddy, K.V.V.S. (2013) 'Text independent speaker identification with finite multivariate generalised Gaussian mixture model and k-means algorithm', Int.

Face Recognition Algorithm based on Doubly Truncated Gaussian Mixture Model using DCT Coefficients

“…The likelihood function contains the number of components M which can be determined from the k-means algorithm given by sailaja et al (21). The k-means algorithm requires the initial number of components which can be taken by plotting the histogram of the face image using MATLAB code and counting the number of peaks.…”

“…Hence in this paper we develop and analyze a face recognition model based on Doubly truncated multivariate GMM. The doubly truncated Gaussian mixture model is capable of portraying several probability distributions like asymmetric / symmetric / platy kurtic / lepty kurtic distributions [19,21]. Various approaches are discussed by different researchers on the problem of face recognition.…”