| Summary: | This thesis introduces a set of new, dynamic units of visual speech which are learnt using computer vision and machine learning techniques. Rather than clustering phoneme labels as is done traditionally, the visible articulators of a speaker are tracked and automatically segmented into short, visually intuitive speech gestures based on the dynamics of the articulators. The segmented gestures are clustered into dynamic visemes, such that movements relating to the same visual function appear within the same cluster. Speech animation can then be generated on any facial model by mapping a phoneme sequence to a sequence of dynamic visemes, and stitching together an example of each viseme in the sequence. Dynamic visemes model coarticulation and maintain the dynamics of the original speech, so simple blending at the concatenation boundaries ensures a smooth transition. The efficacy of dynamic visemes for computer animation is formally evaluated both objectively and subjectively, and compared with traditional phoneme to static lip-pose interpolation.
|
|---|
