This article reports results of a modeling system that is part of a larger project which seeks to achieve perceptually realistic animations of 3D models driven by speech. This system generates individualized animatable 3D face models driven by "muscles " and compatible with MPEG-4. Our main control but on s n the use of an extensive facial database and statistical techniques to generate plausible facial models from limited data: from a single image using eigenface-like statistical recognition techniques, from descriptions (identikit) or randomly with n the parameters of the population represented in the facial database. An interactive module and a muscle-based animation player are also part of the system.