说话人识别中支持向量机核函数参数优化研究

Research on Kernel Parameter Optimization of Support Vector Machine in Speaker Recognition

在基于SVM的说话人识别系统研究中,如何获得理想的识别率是亟待解决的问题。SVM核函数是众多影响识别率因素中最明显的。该系统提高识别率的技术关键是SVM核函数的选取及其参数优化。为此,在对三种常用核函数的特点进行研究的基础上,利用网格搜索法来分别进行参数优选,通过实际语音的训练和识别验证识别效果。目前优选参数可以实现识别率≥99.9%且识别时间<0.1 s。

In the SVM-based speaker recognition system study, it is important to obtain a ideal recognition rate. SVM kernel function is one of the most critical factors that affect the recognition rate. The key technology improves the system recognition rate is the SVM kernel function and parameter optimization. So, based on the characteristics of three kernel functions, it selects the optimal parameters with the grid search method to achieve the training and recognition of the speaker recognition, The current optimized parameters can achieve recognition rate≥ 99.9% and recognition time 〈0.1 s.