Feature Extraction in Speechreading

Abstract

To solve the problem of  feature extraction in speechreading, several appearance-based feature extraction method are compared and a new improved LDA algorithm is proposed in this paper. In speech or speechreading recognition application, Linear Discriminant Analysis(LDA)  usually choose  syllable、HMM state or other units as class unit. but the feature dimensionality reduction  direction based on this traditional LDA have no direct relations with recognition accuracy,To this problem,  A LDA algorithm based on Object (LDAO) which is  fit  for  isolated words recognition in speechreading is proposed, 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 was also given. All experiments were performed on bimodal database, Experimental results showed  that this algorithm is superior to any other appearance-based feature extraction algorithm  in speechreading. Specifically, LDAO  is better than DCT+LDA about 3%.