From raw images of the lips to articulatory parameters: a viseme-based prediction

This paper presents a method for the extraction of articulatory parameters from direct processing of raw images of the lips. The system architecture is made of three independent parts. First, a new greyscale mouth image is centred and downsampled. Second, the image is aligned and projected onto a basis of artificial images. These images are the eigenvectors computed from a PCA applied on a set of 23 reference lip shapes. Then, a multilinear interpolation predicts articulatory parameters from the image projection coefficients onto the eigenvectors. In addition, the projection coefficients and the predicted parameters were evaluated by an HMMbased visual speech recogniser. Recognition scores obtained with our method are compared to reference scores and discussed.