Effect of Temporal Evolution of Songs on Cepstrum-based Voice Signature in Humpback Whales

Cetacean surveys are important for species' population assessments in ecological projects. Acoustics, in this context, provides an efficient, automatic, non-invasive and convenient way to study and observe vocally active cetaceans. Development of a reliable and unbiased acoustic survey method can be a good replacement for conventional tools like line- transect method. However, realization of this goal is yet far from reality. Motivated by recent advancements in acoustics and computational methods, we propose an automatic recognition system for individuals within same cetacean family. The suggested system relies on development of individuals' database based on characteristics of their vocalizations. Such a methodology, accompanied by acoustics based tracking, localization and cetacean behaviour analysis, promises complete social picture of marine eco-system through non-invasive means. In this context, we present our results on recognition of individual humpback whales based on their vocalization data. Previously we reported results using training and test data from the same song. This work analyzes the classification accuracy of data from different songs, i. e. training and test data sets were recorded in different timings. It helps to analyze effects of temporal song evolution on accuracy of our system.