Audiovisual generation of social attitudes from neutral stimuli

The focus of this study is the generation of expressive audiovisual speech from neutral utterances for 3D virtual actors. Taking into account the segmental and suprasegmental aspects of audiovisual speech, we propose and compare several computational frameworks for the generation of expressive speech and face animation. We notably evaluate a standard framebased conversion approach with two other methods that postulate the existence of global prosodic audiovisual patterns that are characteristic of social attitudes. The proposed approaches are tested on a database of ”Exercises in Style” [1] performed by two semi-professional actors and results are evaluated using crowdsourced perceptual tests. The first test performs a qualitative validation of the animation platform while the second is a comparative study between several expressive speech generation methods. We evaluate how the expressiveness of our audiovisual performances is perceived in comparison to resynthesized original utterances and the outputs of a purely framebased conversion system.