A Comparative Study of 2D and 3D Lip Tracking Methods for AV ASR

Abstract

Over the past two decades, many algorithms have been proposed to detect and track a human face and its facial features. Of particular interest to the Automatic Speech Recognition (ASR) community are algorithms that can track the shape of the lips, as such visual speech input can then be used in an auditory-visual (AV) ASR system to improve the recognition accuracy of traditional audio-only ASR systems, particularly in the presence of acoustic noise. Despite the large number of face and lip tracking algorithms that have been proposed over the years, there is a lack of a comparative study that evaluates such algorithms in the context of AV ASR performance. In this paper, the performance of various 2D and 3D lip tracking algorithms is compared from a point of view of AV ASR. In particular, the focus of this study is on algorithms that use explicit lip models. A number of variants of the recently popular Active Appearance Models (AAMs) are compared with a 3D lip tracking algorithm that uses stereo vision. All performance evaluations are made using the AVOZES data corpus.