优先度排序RBF神经网络在与文本无关说话人确认中的应用

TEXT-INDEPENDENT SPEAKER VERIFICATION USING PRIORITY ORDERED RADIAL BASIS FUNCTION NETWORKS

该文介绍了优先度排序径向基函数(PORBF)神经网络的结构与算法,并提出了将其应用于与文本无关说话人确认时的训练算法、似然度的计算方法以及识别规则。为了增强PORBF网络的泛化能力,该文用压缩矢量构造抑制样本集,提出了顺序选取、最近邻选取和最远距离选取等3种选择抑制样本集中说话人的方法,并对PORBF神经元的输出进行了等比递减加权.在相同条件下的与文本无关说话人确认实验中,传统的矢量量化方法的等差错率可达10.56％,而基于PORBF网络的确认系统使用最近邻选择方法构造抑制样本集,其等差错率可达6.83％;性能提高很多。

The structure and algorithm of Priority Ordered Radial Basis Function (PORBF) Networks is introduced. The concrete training algorithm, calculational methods of likelihood score and verification rule, used for text-independent speaker verification, are proposed. To enhance the generalization ability, the compressing vectors are applied to construct the inhibitory samples set and three methods including sequential selection, nearest neighbor selection and furthest distance selection are presented for the choose of anti-speakers. Moreover, the outputs of neurons are weighted by a descendent array. Using these algorithms and methods, the performance is examined by a series of experiments. The results show that under the identical experiment conditions, when the inhibitory set is composed of anti-speakers' compressing vectors selected using nearest neighbor method, the Equal Error Rate (EER) using PORBF networks can decreased to 6.83% from 10.56% using conventional VQ method. For speaker verification, the PORBF network provides better performance than the VQ classifier.