This work focuses on the study of voice production. The main objective is to propose solutions for its analysis-synthesis which enable a wide representation of the diversity of biometric signals (observed in practice) within the framework of source-filtre paradigm or simplified physical models. The study is conducted according to three axes: exploration of laryngeal activity, modelisation of the phonation and sound analysis-synthesis. After a description of the vocal apparatus and current analysis-synthesis models, the first part focuses on the exploration of biometric signals. From an inventory of current devices, we propose a synchronisation tool which reveals the complexity ot the 3D motion of vocal folds. A comparative study with a similar human-valve-based sound production system (lips of a tombonist) is conducted and the development of a new tool combining two recent electrical imaging techniques is proposed. In the second part, two new analysis algorithms are proposed for a source-filter model based on a causal-anticausal all-pole representation of speech signals. A computation method of the complex cepstrum is proposed, which ensures the holomorphy on a convergence domain in the Z plane. This tool is then used for the separation of the source and the filter. Finally, parametric representations of glottal area waveforms are used to drive models controlled by the geometry. A synthesis model, based on the source-filter model but including the effect of the coupling between the glottis and the vocal tract, and advances in the modeling of the vocal tract with regular profile are presented.
