Lip reading using DWT and LSDA

In lip reading, selection of feature play crucial role. Goal of this work is to compare the common feature extraction modules. Proposed two stage feature extraction technique is exceedingly discriminative, precised and computation efficient. We have used, Discrete Wavelet Transform (DWT) to decorrelate spectral information and extract only the salient visual speech information from lip portion. In the second stage the Locality Sensitive Discriminant Analysis (LSDA) is used to further trim down the feature dimension while preserving the required identifiable ability. A competent feature extraction module result a novel automatic lip reading system. We have compared performance of classical Naive Bayes with the popular SVM classifier. The CUAVE database is used for experimentation and performance comparison. Experimental results show that DWT+LSDA feature mining is better than DWT with PCA or LDA. The performance of Naïve Bayes classifier is exceedingly augmented with DWT+LSDA.