摘要
传统的基于最大似然估计高斯混合模型参数的方法是一种无导师的学习方法，该方法的主要缺点是学习算法在估计一类模式模型中的参数时只利用了该类模式中的训练样本，而未考虑其它类训练样本的分布影响，因此，这种方法的识别效果往往不够理想．针对以上问题，作者提出利用最小误分率估计高斯混合模型参数的方法，这种方法考虑了不同类之间的样本的区分性．同时为了提高获得全局最优解的可能性，文中给出一种利用遗传规划求解最优参数的算法．这种方法用于非限定文本的话者识别．实验表明，该方法较传统的参数估计方法识别效果好．
The traditional approach for estimating parameters in Gaussian Mixture Models(GMM) based on maximum likelihood is a kind of unsupervised learning method, its shortage is that the parameters in GMM are derived only by the training samples in one class without taking into account the effect of sample distributions of other classes, hence, its recognition is usually not ideal. This paper presents an approach for estimating parameters in GMM based on the minimum classification error rate of different classes, this method takes into account the discriminations of samples in different classes. To increase the possibility of obtaining the global optimal solution, this paper proposes an approach for estimating the optimal parameters in GMM based on Evolutionary Programming. An experiment has been conducted using the method for text independent speaker recognition, the results have shown that the recognition accuracy is higher than that of the traditional approach.
出处
《计算机学报》
EI
CSCD
北大核心
1999年第8期804-808,共5页
Chinese Journal of Computers
基金
国家八六三高技术研究发展计划
国家自然科学基金
关键词
最小误分率
高斯混合模型
模式识别
语音识别
Minimum classification error rate, gaussian mixture model, pattern recognition.
