Recent progresses in animation have allowed the use of virtual character to many extents. Virtual signers (or signing avatars) are three-dimension characters expressing themselves in sign language. These characters allow the the broadcasting of audio information to deaf and hearing-impaired signers in an anonymous and modular fashion. However, automatic generation of animation for such characters strongly rely on the lexical description of signs. Signs described through these models are usually perfect and geometric performances leading to robotic and unrealistic movements. These thesis focuses on adding information to the control skeleton of the signer to help him perform signs in a more human and realistic way. Such information are grouped under the name of "anatomic model" and are divided in five main contributions : a new computer-based description of the skeleton, an anthropometric study of the hand, the merging of articulatory dependencies, a new model of the carpo-metacarpal complex allowing an easier opposition of the thumb and finally a model computing posture comfort. These contributions are then implemented in a generation system through linguistic contraints-adapted techniques. The study ends with an evaluation of the system and a the presentation of future prospects.
