Weakly supervised click models for odontocete species classification

This paper addresses the problem of automatic learning of statistical models of clicks for odontocete species classifications, particularly focusing on improving accuracy of the classifier by iteratively identifying click-like sounds that are likely to be noise and removing these from the model training set. The algorithm is weakly supervised in that no hand-labeled click regions are available, but knowledge of the species present during the time of recording is used. Experiments classifying which of the three species are present show 7-12% reduction in cross species error from a small number of iterations, but also show a need for improved feature extraction to normalize for recording condition bias.