This article describes a new approach to estimate $F_0$ curves using a B-Spline model characterized by a knot sequence and associated control points. The free parameters of the model are the number of knots and their location. The free-knot placement, which is a NP-hard problem, is done using a global MLE within a simulated-annealing strategy. The optimal knots number estimation is provided by MDL methodology. Two criteria are proposed considering control points as real coefficients with variable precision. They differ on the precision used. Experiments are conducted in a speech processing context on a 7000 syllables french corpus. We show that a variable precision criterion gives good results in terms of RMS error (0.42Hz) as well as in terms of B-spline freedom number reduction (63% of the full model).