DEVELOPMENT OF ALGORITHM FOR MULTIPLE SPEAKER RECOGNITION

Abstract

Multiple speaker recognition is a method used to identify and verify the voice of unknown speaker. Its first part is source separation from the mixed speech signal. Blind source separation (BSS) is a technique of separation of the sources from the mixture without having prior information about actual sources. In this thesis two BSS methods are implemented. One is ICA (Independent component analysis) and other is DUET (Degenerate unmixing estimation technique). The experiment is performed in matlab platform for both the methods up to six number of speakers in the mixture and their performance are evaluated and compared by SNR calculated using the actual sources and sources reconstructed after using the above two methods. Independence and WDO assumptions are finally verified and discussed by experimental results using the fast ICA and DUET algorithms. After this second part of speaker recognition is implemented. It involves feature extraction, then speech identification and speech classification. Mel frequency cepstral coefficients are used as feature vectors. Various speech identification techniques and classification techniques such as Vector quantization (VQ), Gaussian mixture modelling (GMM), Hidden markov model (HMM), Gaussian mixture model - Universal background model (GMM—UBM), Support vector machine (SVM) and Artificial neural network (ANN) are used. In ANN, classification using Multiple layer perceptron (MLP), Radial basis probabilistic neural network (RBPNN), Radial basis function network (RBFN), Learning vector quantization (LVQ), Self-organizing map (SOM) have been done. For vector quantization, LBG algorithm and SOM algorithm are used. Finally speech identification and classification accuracy using above methods and their combination are discussed.