Songs of humpback whales (Megaptera novaeangliae) have been studied for several years to gain a deeper insight on the intraspecific social interactions. Such a complex acoustic display is indeed thought to play an important role in both the mating ritual and male to male interaction. Hence, the need to classify the unit constituents of a song objectively and systematically has become crucial to allow processing large data sets. We propose a new approach for song segmentation based on the definition of subunits. Songs of humpback whales collected in Madagascar in August 2008 and 2009 were segmented using an energy detector with a double threshold and classified automatically with a clustering algorithm using MFCCs: the results, which were checked against a manual classification, showed that the use of subunit as the basic constituent of a song rather than the unit produces a more accurate classification of the calls. Such results were expected given that subunits are generally shorter in duration and less variable in terms of their frequency content and so their characteristics are more easily captured by an automatic classifier. Analysis of songs from other years and various areas of the World is necessary to corroborate the repeatability of the method proposed.