van der Schaar, M., Delory, E., Català, A., André, M. Neural network based sperm whale click classification 20th conference of the European Cetacean Society, Gdynia, Poland, Apr 2006 Abstract: Sperm whales often dive in groups for foraging, producing continuous series of clicks. Acoustic recordings of these dives result in a mixture of signals, making it difficult to segment and store the click sequences of individual animals. Since manual separation is an arduous task it would be preferable to automise this process. To this aim a suitable classification function needs to be found, and due to the directional properties of clicks, direct linear classifiers may only work for short sequences and fail when applied to entire dives. Here we study the use of a radial basis function neural network to separate clicks from different whales. Advantages of this type of network are a natural way of processing clustered data and a simple structure allowing fast training through a combination of unsupervised and (linear) supervised techniques. The algorithm is applied to six click series of individually diving males and data containing an entire dive to evaluate the capacity of the algorithm to generalise. It is shown that, depending on the classification parameters, around 90% of the click series can be classified correctly, while for the entire dive this percentage is around 77%. Project: Noise pollution effects on cetaceans