Automatic acoustic estimation of sperm whale size distributions achieved through machine recognition of on-axis clicks.

The waveforms of individual sperm whale clicks often appear as multiple pulses, which are the product of a single pulse reverberating throughout the spermaceti organ. Since there is a relationship between spermaceti organ size and total body size, it is possible to estimate a whale's length by measuring the inter-pulse intervals (IPIs) within its clicks. However, if a click is recorded off-axis, the IPI corresponding to spermaceti organ length is usually obscured. This paper presents an algorithm for automatically estimating the "true" IPIs of sperm whales in a recording by measuring them from on-axis clicks only. The routine works by classifying detected clicks with a support vector machine, assessing the stability of their IPIs, and then clustering the stable IPIs using Gaussian mixture models. Results show that the routine is very accurate in obtaining reliable IPIs, but has a high false negative rate. Nonetheless, since sperm whales click very frequently, it is possible to obtain useful IPI distributions with only a few minutes of recording. This algorithm makes it possible to estimate the body lengths of multiple sperm whales automatically with only one hydrophone. An implementation is available for download at http://whitelab.biology.dal.ca/CABLE/cable.htm.