Most flow models used in numerical simulation of voiced sound production rely, for the sake of simplicity, upon a certain number of assumptions. While most of these assumptions constitute reasonable first approximations, others appear more doubtful. In particular, it is implicitly assumed that the air flow through the glottal channel separates from the walls at a fixed point. Since this assumption appears quite unrealistic, and considering that the position of the separation point is an important parameter in phonation models, in this paper a revised fluid mechanical description of the air flow through the glottis is proposed, in which the separation point is allowed to move. This theoretical model, as well as the assumptions made, are validated using steady- and unsteady-flow measurements combined with flow visualizations. In order to evaluate the effective impact of the revised theory, we then present an application to a simple mechanical model of the vocal cords derived from the classical two-mass model. As expected, implementation of a moving separation point appears to be of great importance for the modeling of glottal signals. It is further shown that the numerical model coupled with a more realistic description of the vocal cord collision can lead to signals surprisingly close to those observed in real speech by inverse filtering.