Initialization in speaker model training based on expectation maximization

摘要

The optimized speaker model is trained by many time iterative algorithm based on expectation maximization (Abbr. EM). In the process, the choice of speaker model initial value has great influence on the final recognition effect. The most common algorithms which are used to choose the initial value are K-means algorithm and LBG algorithm at present, but the two algorithms belong to a sort of local clustering arithmetic, therefore, it is difficult for them to provide the optimal initial value. For this reason, the ant colony algorithm combined with genetic arithmetic is proposed in the paper. The comparative experiment between this algorithm and K-means algorithm has been done, and the experimental results have been obtained to verify that this algorithm can bring better recognition rate than K-means algorithm.