Improved lip contour extraction for visual speech recognition

Abstract

Automatic speech recognition systems perform on acoustic speech signals and therefore they are unreliable in noisy environments. Visual speech features such as lip movements of a speaker can make the speech recognition system robust. To track the lip movements, lip contour extraction is a necessary step and plays a crucial role in the visual speech recognition. In this paper, we propose a new method for lip contour extraction using fuzzy clustering with elliptic shape information and active contour model. In this method, we combined both image and model based methods to improve the performance of lip contour extraction. Our proposed lip contour extraction method outperforms few of existing lip contour extraction methods. We applied our lip contour extraction method on 3600 lip images from VidTimit database and results are found better than the few existing lip contour extraction methods.