Feature extraction based on LDAO algorithm in speechreading

In speech or speechreading recognition application, traditional LDA algorithm usually choose syllable, HMM state or other units as class unit. but the feature dimensionality reduction direction based on this traditional LDA has no direct relations with recognition accuracy. To this problem, An improved LDA algorithm based on Object (LDAO) which is fit for isolated words recognition in speechreading is proposed in this paper, LDAO choose the objects to be recognized as class unit to Linear Discriminant Analysis, which guarantees feature extraction follow the most discriminant directions among objects in theory. Subsequently, training and recognizing method for LDAO are also given. Experimental results on bimodel database showed that this algorithm is better than traditional LDA. Specifically, LDAO is better than DCT+LDA about 3%.