In this paper we present some experiments on multiclass ToBI pitch accent classification. The system is based on the fusion of pairwise classifiers, which are specialized in the distinction of pairs of prosodic labels. Several machine learning techniques, including neural networks, decision trees and support vector machines, are combined in different ways in order to find the best overall combination. Variations of pair-wise classifiers are introduced in order to take into account the influence of the samples of the remaining classes during the training of the binary classifiers. The use of these techniques allowed us to improve the results, both the overall classification accuracy and the balance across the different ToBI pitch accent classes.