Phys. Rev. E 93, 022138 (2016)

Credit: © JAN BAKS / ALAMY STOCK PHOTO

Monitoring the fascinating sounds made by whales — or any other species of animal in the wild, for that matter — is made challenging by the fact that recordings typically lack information regarding the sender and its context. And given the different sensory-processing systems of these animals, is it even meaningful to categorize single calls according to features that seem relevant to human observers?

Heike Vester and colleagues have opted for a completely different approach. Instead of separating and sorting vocalizations into types, they compared ensembles of sounds produced by different groups of long-finned pilot whales recorded in northern Norway. They achieved this by performing a cepstral decomposition of the noise signal, a process akin to a spectral decomposition. Computing and analysing the distribution of the resulting cepstral coefficients allowed the authors to identify different whale groups in a statistically significant way.

Taking their cue from machine learning, where ensembles are sometimes referred to as bags, Vester and co-workers termed this the bags-of-calls-coefficients approach. One might say the results of this approach speak for themselves: the authors uncovered differences between social groups of whales, consistent with the existence of distinct vocal cultures and dialects among the cetaceans.