中文摘要：

在语音与唇读识别应用中,传统的LDA（linear discriminant analysis）算法一般以音节、半音节、HMM状态等基元为类别进行数据分段,经线性判别分析后获得的特征投影方向与识别率不直接相关,影响了识别率。提出了一种新的基于LDAO（linear discriminant analysis based on object）的唇读特征提取算法,该算法以待识别对象为类别进行线性判别分析,在理论上保证了唇读特征矢量向最具判别能力的方向投影。基于唇读数据库的实验证明,该算法明显优于现有各种唇读特征提取算法,比DCT＋LDA算法识别率提高了3%。

英文摘要：

In speech and lipreading recognition application,LDA（linear discriminant analysis）algorithm is usually based on syllable,semi-syllable,HMM state or other class units.But the extracted features based on traditional LDA have no direct relation to recognition accuracy.This paper proposed linear discriminant analysis based on object（LDAO） algorithm on recognizing isolated words in lipreading.It selected objects to be recognized as class to LDA,which ensured feature extracting followed the most discriminant directions among objects in theory.Experiments on bimodal database show that this algorithm is superior to any other feature extracting algorithms in lipreading.Specifically,the recognition accuracy is better than DCT＋LDA algorithm about 3%.