Support Vector Machine (SVM) is a powerful tool for binary classification. Numerous methods are known t o fuse several binary SVMs into multi-class (MC) classifiers. These methods are efficient, but an ac curate study of the misclassified items leads to notice tw o kinds of error which could be avoided: (1) Some items are undetermined because the decision process does not use the entire information from the SVM, and (2) some undetermined items are not processed in the fashion which would suit them. In this paper, we present a method which partially improves these two points by applying some result of Belief Theories (BTs) to SV M combination, while keeping the efficient aspect of the classical methods.