This paper introduces an unsupervised model for melodic segmentation that extends a method initially proposed in computational biology. In the model segments are identified as sections of maximal contrast within a musical piece, using for this the Jensen-Shannon divergence. The model is extended upon its original formulation, and experiments to test its performance are carried out for a small set of selected folk song melodies. Generalization of the model is tested on 100 folk songs. Our results show a significant improvement upon the model's original formulation. In addition, we situate our model in the context of a cognition-based ensemble learning framework and justify its use within it. The need for such a cognition-based ensemble approach is also discussed.