改进的Fisher准则及其在语音聚类中的应用

针对传统的Fisher准则在聚类分析方面的不足，提出了一种新的修正的Fisher准则。通过重新定义类内方差，得出最大Fisher值与最佳聚类数目的对应关系，从而预测出最佳聚类数目。通过对语音库中语音段落的说话者组成结构进行分析，证明了修正后的Fisher准则能有效地优化聚类分析，并显著提高了说话者组成结构分析结果的精度。应用基于修正后的Fisher准则的LBG聚类算法，能为语音操控系统中无监督的说话人自适应提供更可靠的说话人模板库。

An improved Fisher criterion for effective cluster optimization

In order to overcome the drawback of the conventional Fisher criterion on cluster analysis, an improved Fisher criterion is developed in this paper. Through redefining the within-class scatter variance, the corresponding relation of the maximum of Fisher value and the best cluster number is defined, so the best cluster number is forecasted. After experimenting on the speech database, it is proved that the improved Fisher criterion can optimize the cluster analysis effectively and improve the precision of speakers composition structure analysis result evidently. More reliable speaker template database can be provided for unsupervised speaker adaptation in voice-operated control system using LBG clustering based on modified Fisher criterion.