The theory of belief functions has been successfully used in many classification tasks. It is especially useful when combining multiple classifiers and when dealing with high uncertainty. Many classification approaches such as k-nearest neighbors, neural network or decision trees have been formulated with belief functions. In this paper, we propose an evidential calibration method that transforms the output of a classifier into a belief function. The calibration, which is based on logistic regression, is computed from a likelihood-based belief function. The uncertainty of the calibration step depends on the number of training samples and is encoded within a belief function. We apply our method to the calibration and combination of several SVM classifiers trained with different amounts of data.