摘要
提出一种新的可用于说话人识别的径向基函数网络(RBFN)阵列.RBFN网设计思想是在确定网络中心点之后采用最小线性方差作为目标函数解得最优权重,该方法并不能得到最优分类效果.使用Fisher目标函数,替代RBF中的误差目标函数来求取最优权重,用与文本无关的闭集说话人识别系统对该算法进行了验证,实验结果表明,该方法提高了RBF分类能力,比传统的RBF算法以及ROLS算法具有更高的识别率,并在识别效果接近GMM方法的情况下计算量大幅度减少.
A novel RBF networks array for speaker recognition is proposed. Usually in design of RBF networks, after the centers have been fixed, the optimal weights are determined by minimizing least squares function.However this method can not lead to optimal classification. A new RBF designing method based on Fisher discriminant function is presented. Experiments on a closed set, text-independent speaker recognition system show that higher accuracy can be achieved through Fisher RBF networks than classical RBFN and ROLS algorithm. Moreover,this method is more efficient than GMM.
出处
《复旦学报（自然科学版）》
CAS
CSCD
北大核心
2005年第1期118-121,127,共5页
Journal of Fudan University：Natural Science
基金
国家自然科学基金资助项目(60171036
30370392)
