Machine Recognition and Applications

Abstract

Introductory comments by Juergen Luettin. He mentions earlier discussions indicated that we can’t really clearly determine which features are important for speechreading. Most important features might be the lip contours, other features could be the visibility of the tongue and the teeth. He suggests that visual features could be divided into acoustic and linguistic features, e.g. features like eye brows might provide higher level linguistic information rather than acoustic information. If we don’t know which features are important we could consider using the whole face image. But then our system has to learn the variances which account for different speech unit and those which account for linguistic variability and image variability (lighting, scale ...). If, on the other hand we only use certain features, we can significantly reduce the learning problem but we might disregard important features.