Improved visual speech synthesis using dynamic viseme k-means clustering and decision trees

We present a decision tree-based viseme clustering technique that allows visual speech synthesis after training on a small dataset of phonetically-annotated audiovisual speech. The decision trees allow improved viseme grouping by incorporating k-means clustering into the training algorithm. The use of overlapping dynamic visemes, defined by tri-phone time-varying oral pose boundaries, allows improved modelling of coarticulation effects. We show that our approach leads to a clear improvement over a comparable baseline in perceptual tests. The avatar is based on the freely available MakeHuman and Blender software components.