摘要
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。
By using the K-means and the K-nearest neighbor heuristic in estimating the elliptical basis function (EBF) parameters may produce sub-optimal performance when the input vectors contain correlated components. And how to determine an appropriate number of clusters of EBFN is also a problem. This paper proposes to overcome those problems by incorporating the improved RPCL (rival penalized competitive learning) algorithm and the EM (expectation maximization) algorithm into the EBF structure. The speaker verification experiment result shows this method has a better representation of data set and a lower verification error rate.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2003年第2期204-207,共4页
Pattern Recognition and Artificial Intelligence
基金
上海市曙光学者计划资助项目
关键词
EBF网络
RPCL算法
EM算法
参数
话者识别
语音识别
Elliptical Basis Function (EBF) Networks, Rival Penalized Competitive Lerning, Expectation Maximization (EM) Algorithm