通常我们用K-平均法和K-邻近法估计椭圆基函数(EBF)中心位置与函数宽度等参数。但上述方法在输入矢量包含相关元素时存在性能次优化问题。另外,对于EBF网络来说,如何选择适当的类的数目仍是一个难以解决的问题。本文提出用结合改进的RPCL算法和EM算法的EBF网络结构来解决上述问题。在话者识别实验中,证明这种结构具有更优越的样本表征能力以及更好的识别率。 Usually we use K-average method and K-adjacent method to estimate the ellipse basis function (EBF) center position and function width and other parameters. However, the above method has the problem of sub-optimal performance when the input vector contains the relevant elements. In addition, how to choose the appropriate number of classes for the EBF network remains a difficult issue to solve. This paper proposes to solve the above problems by using the EBF network structure combining the improved RPCL algorithm with the EM algorithm. In the speaker recognition experiment, this structure is proved to have superior ability of sample characterization and better recognition rate.