A New Genetically Optimized GMM for Speaker Recognition

The traditional training method of Gaussian mixture model is sensitive to the initial model parameters, and easy to lead to a sub-optimal model in practice. To resolve this problem, it utilized the niche hybrid genetic algorithms (NHGA) to find the optimum model parameters. It provided a new architecture of hybrid algorithms, which organically merged the niche techniques and maximum likelihood (ML) algorithm into GA. It used the niche techniques to make the exploration capabilities of GA stronger, and the ML algorithm to make the exploitation capabilities of GA more powerful. Besides, it used a heuristic updating strategy to control the GA mixture crossover rate P
c
and mutation rate P
m
. Experiments were based on an independent speaker recognition system. The results from PKU-SRSC database show that this method can obtain more optimum GMM parameters and better results than the traditional and the improved GMM for speaker recognition.