Performance Analysis of Speaker Identification System Using GMM with VQ

Personal identity identification is an important requirement for controlling access to protected resources. Biometric identification by using certain features of a person is a more secured solution for security identification. Advances in speech processing technology and digital signal processors have made possible the design of high-performance and practical speaker recognition systems. A more flexible speaker identification system is able to operate without explicit user cooperation and independency of the spoken utterance (textindependent mode).This paper proposes a system for text independent speaker identification by extracting MFCC features and implementing optimized GMM speaker modeling. Expectation Maximization algorithm is used to compute the GMM parameters. Performance of the proposed system is evaluated based on its identification accuracy.It is compared with the system using VQ speaker modeling technique. A TIMIT database of 100 speakers is used to study the performance of the proposed system. Key terms: Feature extraction, Speaker modeling, vector quantization, speaker identification, Mel-frequency cepstral coefficients(MFCC), Gaussian mixture model(GMM),Gaussian mixture model-Expectation maximization(GMM-EM)