Three-dimensional facial adaptation for MPEG-4 talking heads

Abstract

This paper studies a new method for three-dimensional (3D) facial model adaptation and its integration into a text-to-speech (TTS) system. The 3D facial adaptation requires a set of two orthogonal views of the user's face with a number of feature points located on both views. Based on the correspondences of the feature points' positions, a generic face model is deformed nonrigidly treating every facial part as a separate entity. A cylindrical texture map is then built from the two image views. The generated head models are compared to corresponding models obtained by the commonly used adaptation method that utilizes 3D radial bases functions. The generated 3D models are integrated into a talking head system, which consists of two distinct parts: a multilingual text-to-speech sub-system and an MPEG-4 compliant facial animation sub-syslem. Support for the Greek language has been added, while preserving lip and speech synchronization.