Abstract: To solve the key problem of extracting visual speech feature in lipreading，a method based on DCT and LDA is proposed.To extract most discriminative visual feature among different mouth classes，first，DCT is performed on the visual speech region；and then based on LDA the most discriminative feature vector is extracted from DCT coefficients.The experiments on speaker-dependent，speaker-independent database and in real-time lipreading environment show that this method is more effective than traditional manual DCT coefficients extraction method and PCA feature extraction method.

Key words: lipreading, feature extraction, Discrete Cosine Transformation（DCT）, Linear Discriminative Analysis（LDA）

摘要： 为解决视觉语言特征提取这个唇读技术中最关键的难题，提出一种新的基于DCT和LDA的特征提取方法。为提取对不同口型最具分类能力的特征矢量，首先基于DCT对视觉语言部位变换降维，然后基于LDA算法从DCT系数提取对口型分类性能最优的特征矢量。在特定人与非特定人的唇读数据库上以及实时唇读识别的实验都表明，该方法唇读识别率比传统的人工直接选择DCT系数法以及PCA提取法有明显提高。

关键词: 唇读, 特征提取, 离散余弦变换（DCT）, 线性判别分析（LDA）