Abstract

We describe a method to automatically classify Humpback whale (Megaptera Novaeangliae) song that offers improvements over matched spectrogram techniques currently widely employed. Humpback song is a useful training example for a range of ocean acoustic transient detection and classification problems because it consists of units of varying length, frequency range and type, from nearly tonal to highly transient. With any recognition system it is vital that the data is first segmented into appropriate units. This is nontrivial and often implemented manually. We have developed a segmentation using wavelet packet decompositions that also produces a 'feature vector' with which to classify the data using a neural network. The next step is to select the network architecture, where there are many good alternatives, including a principle component front end coupled to a back-propagation network and self-organising networks with Learning Vector Quantisation. Various architectures typically achieve 80% classification rates on a challenging variety of units. The approach has the added benefits of being shift invariant with respect to time, and somewhat tolerant of frequency and time stretching. Since the methods employed are not specific to whale song the approach can be usefully applied to other types of marine transient signals with minimum modification.