Evaluation de systèmes de génération de mouvements faciaux
Oxana Govokhina, Gérard Bailly, Gaspard Breton, Paul Bagshaw.

This paper presents the implementation and evaluation of different movement generation techniques for speech-related facial movements. State-of-the-art systems are implemented. A novel system that combines HMM-driven pre-selection of diphones with a standard concatenation system is also implemented. The trajectory formation systems are parameterised using the same training material. The groundtruth data consists of facial motion and acoustic signals of one female speaker uttering 238 sentences. Both objective and subjective evaluation of the systems is reported. The objective evaluation observes the linear correlation coefficient between original and predicted movements. It is complemented by an audiovisual preference test where ground-truth and predicted movements drive a 3D virtual clone of the original speaker.