The science behind Howl Coder

The Canid Howl Project is the work of a large number of scientists, trying to understand the range of different vocal behaviours of canids: primarily wolves, dogs, and coyotes. Despite being closely related, these animals have very different ways of communicating with howls, barks, yips, growls, etc. We hope to understand more about the whole range of canid species and breeds, by studing their vocal behaviour.


Our ears are not evolved to pick up the subtleties of the howls and other sounds of wolves, dogs, and coyotes. So we use a mathematical technique called the Fourier transform to identify the different frequencies (pitches) present at any point in time. Applying a Fourier transform repeatedly produces a spectrogram, which shows time along the horizontal axis, and frequency along the vertical axis. Stronger frequencies are shown as brighter colours. For example, in the spectrogram shown below on the left, there is one strong frequency that remains fairly constant throughout the recording. In contrast, in the howl on the right, the pitch goes up and down.


 

We want to look at exactly how the animals are varying the pitch of their howls with time, and relative to other animals that are howling at the same time (chorus howling). This can give us information on the meaning of the howls (territory defence, preparation for hunting, etc), as well as the individual identity of the howler, and possibly also the identity of the pack. To do this, we need to measure all the pitch variations that appear on the spectrogram. It might seem that this would be easy to do automatically using a computer, but it's not so easy! Take a look at the spectrogram below on the left. Multiple animals are howling, overlapping each other in time and frequency; the lines are crossing and it's not even clear where the continuation of a howl goes after a crossing. Automatic algorithms can work for some species, like the image below on the right, which shows the analysis of a recording of a pilot whale. But these canid recordings are too noisy, and too complex to analyse automatically.


 

Humans are especially skilled at finding patterns in pictures. No computer algorithm can match the performance of the human brain at this task. That's why we're turning to you - hundreds or thousands of humans - to do what computers cannot. Track the howls on the spectrogram, and we can convert the lines you draw back into time and frequency, and use this to find out exactly what sounds the animals are making.


But what happens when people make a mistake? Or if they don't agree on the interpretation of the spectrogram? There are statistical techniques that can be used to take this into account; eliminating obviously incorrect scribbles, and taking the majority decision where there is disagreement. So don't worry, just do your best, and the maths will straighten everything out in the end!


Top of page -->