Chee-Ming Ting Sh-Hussain Salleh Tian-Swee Tan A. K. Ariff. International Conference on Intelligent and Advanced Systems 2007 Jain-De,Lee INTRODUCTION GMM SPEAKER IDENTIFICATION SYSTEM EXPERIMENTAL EVALUATION CONCLUSION Speaker recognition is generally divided into two tasks ◦ Speaker Verification(SV) ◦ Speaker Identification(SI) Speaker model ◦ Text-dependent(TD) ◦ Text-independent(TI) Many approaches have been proposed for TI speaker recognition ◦ VQ based method ◦ Hidden Markov Models ◦ Gaussian Mixture Model VQ based method Hidden Markov Models ◦ State Probability ◦ Transition Probability Classify acoustic events corresponding to HMM states to characterize each speaker in TI task TI performance is unaffected by discarding transition probabilities in HMM models Gaussian Mixture Model ◦ Corresponds to a single state continuous ergodic HMM ◦ Discarding the transition probabilities in the HMM models The use of GMM for speaker identity modeling ◦ The Gaussian components represent some general speakerdependent spectral shapes ◦ The capability of Gaussian mixture to model arbitrary densities The GMM speaker identification system consists of the following elements ◦ Speech processing ◦ Gaussian mixture model ◦ Parameter estimation ◦ Identification The Mel-scale frequency cepstral coefficients (MFCC) extraction is used in front-end processing Input Speech Signal FFT Pre-Emphasis Triangular band-pass filter Frame Logarithm Mel-sca1e cepstral feature analysis Hamming Window DCT The Gaussian model is a weighted linear combination of M uni-model Gaussian component densities M p( x | ) wi bi ( x ) i 1 Where x is a D-dimensional vector bi ( x), i 1,...,M are the component densities wi , i=1,…,M are the mixture weights M The mixture weight satisfy the constraint that w i 1 i 1 Each component density is a D-variate Gaussian function of the form bi( x ) 1 T 1 exp{ ( x i ) i ( x i )} D/2 1/ 2 (2 ) | i | 2 1 Where i is mean vector i is covariance matrix The Gaussian mixture density model are denoted as (wi , i , i ),i 1,...,M Conventional GMM training process Input training vector LBG algorithm N EM algorithm Convergence End Y Input training vector Overall average N Y m<M Split D’=D Clustering Cluster’s average Calculate Distortion N (D-D’)/D< δ Y End Speaker model training is to estimate the GMM parameters via maximum likelihood (ML) estimation p( X | ) p( xt | ) T t 1 Expectation-maximization (EM) algorithm 1 T wi p(i | xt , ) T t 1 p ( i | x , ) x t t i t T1 p ( i | x t 1 t , ) T 2 p ( i | x , ) x t t 2 i2 t T1 i p ( i | x t 1 t , ) T This paper proposes an algorithm consists of two steps Cluster the training vectors to the mixture component with the highest likelihood Ci arg maxbi ( x ) 1i M Re-estimate parameters of each component wi number of vectors classified in cluster i / total number of training vectors i sample mean of vectors classified in cluster i. i sample covariance matrix of vectors classified in cluster i The feature is classified to the speaker Sˆ ,whose model likelihood is the highest Sˆ arg max p( X | k ) 1k S The above can be formulated in logarithmic term ˆ S arg max log p( xt | k ) T 1 k S t 1 Database and Experiment Conditions ◦ 7 male and 3 female ◦ The same 40 sentences utterances with different text ◦ The average sentences duration is approximately 3.5 s Performance Comparison between EM and Highest Mixture Likelihood Clustering Training ◦ The number of Gaussian components 16 ◦ 16 dimensional MFCCs ◦ 20 utterances is used for training Convergence condition | p( X | (k 1) ) p( X | (k ) ) | 0.03 The comparison between EM and highest likelihood clustering training on identification rate ◦ ◦ ◦ ◦ 10 sentences were used for training 25 sentences were used for testing 4 Gaussian components 8 iterations Effect of Different Number of Gaussian Mixture Components and Amount of Training Data ◦ MFCCs feature dimension is fixed to 12 ◦ 25 sentences is used for testing Effect of Feature Set on Performance for Different Number of Gaussian Mixture Components ◦ Combination with first and second order difference coefficients was tested ◦ 10 sentences is used for training ◦ 30 sentences is used for testing Comparably to conventional EM training but with less computational time First order difference coefficients is sufficient to capture the transitional information with reasonable dimensional complexity The 12 dimensional 16 order GMM and using 5 training sentences achieved 98.4% identification rate