A New Gaussian Mixture Model Optimization Method

The traditional training methods of Gaussian mixture model (GMM) are sensitive to the initial parameters, and when the training data is limited, it has weak generalization. To resolve theses problems, it proposed a novel GMM optimization method. It used the fuzzy expectation maximization approach and the niche technique to form new hybrid architecture, which can reduce the possibility of premature convergence presence and improve the exploitation capabilities of genetic algorithms (GA). To increase the accuracy of classification and make GMM more generalized, the other people´s discriminative information was integrated into fitness function. Besides, it also used an adaptive updating strategy to control the GA parameters. The experimental results show this method can obtain more optimum GMM parameters and better results than the traditional and the two improved versions for speaker recognition.