Abstract : The sinusoidal model of speech signals is usually defined on a “short-term” basis, i.e. on successive frames of about 10–30 ms. In this thesis, we add to this usual spectral modeling a new level of modeling along the temporal axis: the goal is to model the temporal trajectories of the sinusoidal parameters (amplitudes and phases) over durations which are significantly longer than the short-term frames (typically several hundreds of ms; continuously voiced sections of speech are considered in this study). For this, we propose to use different long-term models based on discrete cosine and polynomial functions. The fitting of these models with the parameters trajectories is achieved by a weighted least square minimisation technique, the weights being derived from perceptual criteria which are adapted to the long-term processing. For this task, a series of iterative algorithms is proposed and tested. The proposed long-term approach is shown to provide an efficent and sparse representation of the dynamics of voiced speech signals.