5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

A Hierarchy Probability-Based Visual Features Extraction Method for Speechreading

Yanjun Xu, Limin Du, Guoqiang Li, Ziqiang Hou

Inst Acoustics, Chinese Acad Sci, China

Visual feature extraction method now becomes the key technique in automatic speechreading systems. However it still remains a difficult problem due to large inter-person and intra-person appearance variabilities. In this paper, we extend the normal active shape model to a hierarchy probability-based framework, which can model a complex shape, such as human face. It decomposes the complex shape into two layers: the global shape including the position, scale and rotation of local shapes (such as eyes, nose, mouth and chin); the local simple shape in normal form. The two layers describe the global variation and local variation respectively, and are combined into a probability framework. It can perform fully automatic facial features locating in speechreading, or face recognition.

Full Paper

Bibliographic reference.  Xu, Yanjun / Du, Limin / Li, Guoqiang / Hou, Ziqiang (1998): "A hierarchy probability-based visual features extraction method for speechreading", In ICSLP-1998, paper 0187.