The songs produced by male humpback whales during the breeding season have been the interest of researchers for a long time because of their complexity and the curiosity they arise in the general public.
The songs produced by male humpback whales during the breeding season have been the interest of researchers for a long time because of their complexity and the curiosity they arise in the general public. The aim of this study was to develop a method for automatic classification of humpback whale Megaptera novaeangliae songs into their unit components - defined by Payne as continuous sounds between two silences. The recordings were collected in August 2007 and 2008 in the channel between the east coast of Madagascar and the Island of Sainte Marie, where humpbacks gather from June to October for breeding purposes. The analysis was carried out using an energy approach for the segmentation of a song where two thresholds, i.e. threshold of start and threshold of end, were used to identify the location of each vocalisation. In addition, the latter were classified using the k-means method and a comparison was drawn between the classification obtained using predictors which are commonly employed for speech analysis. Namely, these are the linear prediction coefficients (LPCs) and Mel-frequency cepstrum coefficients (MFCCs) and the cepstrum coefficients. The validity of the clustering obtained was then tested by manually conducting a spectrographic analysis of the vocalizations to see if similar units were classified in the same group or not. The issues encountered with the processing of the songs using the segmentation algorithm were mainly derived by the difficulties encountered to take good recordings in the field. In addition, detailed analysis of the vocalisations showed that the features of a unit can change abruptly throughout its duration making it difficult to characterise and cluster them systematically. For this reason, the focus of research is now aimed at developing a new approach for song segmentation based on the identification of subunits that are characterised by looking at the changes of their frequency content through time. The distinction between subunits and units should improve the accuracy of classification algorithms, especially for those vocalisations that present a complex structure that varies significantly with time.