摘要
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.
出处
《信号处理》
CSCD
2002年第5期442-447,共6页
Journal of Signal Processing
基金
上海市重点学科建设项目资助
关键词
互信息测度
聚类特性
语音识别
距离测度
Cluster property Mutual information measure Speech recognition Distance measure