Visual speech understanding using independent component analysis

The purpose of this work is to enhance the performance of visual speech recognition by using independent component analysis (ICA) in order to extract statistically independent visual features. In the first place, we derive the optical flow fields for consecutive frames of people speaking. Then, we use ICA in order to derive the basis images for these optical flow fields. The coefficients of these basis flow fields will comprise the visual features of interest. We will show that using ICA on optical flow fields yields better classification results than the traditional approaches based on principal component analysis (PCA) for instance. Our approach is evaluated for the Tulipsi database and compared to the standard approaches.