语音识别中互信息测度的聚类特性分析与实验评价

Cluster Property Analysis and Experimental Evaluation of Mutual Information Measure for Speech Recognition

互信息作为一种模式距离测度已经被成功地应用在语音识别中,并由此提出了语音识别的互信息匹配模型。本文运用统计方法对互信息测度的聚类特性进行了分析,对其实际识别性能进行了实验评价,并与传统的距离测度Euclidean,Mahalanobis和Itakura-Saito进行了比较。分析与实验表明,互信息测度具有较好的聚类特性,相应的类内凝聚度较高,类间耦合度较小,在采用线性预测倒谱系数LPCC作为特征参数时,运用互信息测度的错误识别率较小,仅为运用Euclidean测度时的50％。

Mutual information as a measure of pattern difference has been successfully used in speech recognition, which forms a basis of the so-called mutual-information matching model. In this paper, cluster properties of the mutual information measure are studied based on a statistical analysis and comparison with some traditional measures including Euclidean, Mahalanobis and Itakura-Saito. Recognition performance is evaluated, and comparisons with other measures are made. The statistical analysis and experiments show that the mutual information measure has clear preponderance over the other three measures.