How animals stay together through sound

How do animals coordinate with one another?

Have you ever tried to coordinate with a group? Maybe you and your friends were trying to decide where to go to dinner, or even what route to take when walking to the restaurant. To achieve coordination in these cases, you probably had to communicate with your friends. And more often than not, you probably communicated using sound, i.e. vocally.

When we look across nature, we see many examples of group coordination in non-human animals. From moving synchronously as a group, to banding together against common threats, coordination is often a part of social living. And just as in humans, vocal communication is often crucial in enabling coordination.

In the Communication & Coordination Across Scales project and the CoCoMo group, we are investigating the role that vocal communication plays in allowing groups to coordinate with one another, across different species and different contexts. For example, some social animals have distinct calls that they use to initiate group movement, to stay together while moving, or to recruit others to join them in collective defense against predators. How do the vocal interactions of animals relate to their subsequent behaviors and decisions, and how are these vocalizations used to achieve coordination at a group level?

How do we study coordination in the wild?

Studying group coordination in the wild can be a challenge. We need to capture information on what all members of a social group are doing simultaneously, which is almost impossible to do by observing the animals directly. Luckily, technology can help. 

In our work we typically use bio-logging technology – small tags containing various sensors that can be deployed on animals in the wild to monitor them remotely. Our bio-logging tags usually contain several sensors including GPS, accelerometer, and of course audio. By deploying these tags (temporarily) on all members of a social group, we can get a detailed picture of where all the members of an animal group are moving, what they are doing, and what they are “saying” to one another over time.

The photo below shows a group of meerkats sporting some of our multi-sensor collars. These meerkats are part of the Kalahari Meerkat Project, a long-term study of meerkats in the Northern Cape, South Africa. Collaborating with long-term studies such as the KMP and the Mara Hyena Project (a study of spotted hyenas in the Masai Mara, Kenya) is essential to our work because it allows us to incorporate information on demographics, social structure, and other long-term behavioral information into our studies.

Where does AI come in?

Collecting data on full groups means collecting hours and hours of audio data. To understand how vocalizations are used to coordinate, we first need to comb through these audio recordings to find all of the vocalizations that the animals produce, classify them into types, and figure out which animal produced each call. Unless we’d like to spend the rest of our lives listening to recordings and identifying calls – a task that would drive even the most motivated bioacoustician mad – we need an automated solution. 

Naturally, we turn to machine learning. It turns out, though, that this task isn’t so simple. Unfortunately, animals aren’t inclined to speak clearly into microphones. Instead, they prefer to crash through vegetation, dig noisily in the ground, or loudly crunch their food – all while producing the calls that we are trying to detect! This results in a lot of sporadic noise overlapping their vocalizations, which is hard to filter out. Another problem is that their calls are fairly infrequent. Even for a very vocal animal such as a meerkat, there is far more time when they are not calling compared to when they are, leading to a needle-in-a-haystack problem.

Here’s an example of some meerkat audio data, shown as a waveform (top) and spectrogram (bottom). The broadband noise (reddish vertical stripes) is the sound of meerkats digging, and the regions marked with dashed lines are the calls. You can listen to the clip here:

How we solve these tricky problems is a story for another day, but if you’re interested you can check out our self-supervised learning approach, called animal2vec, here.

Once we have detected and classified the calls, we can finally bring everything together! Below is a visualization of a group of 7 meerkats as they go about their day foraging for food, moving and calling. The filled dots represent the meerkats, and the markers popping up around the dots represent the vocalizations that each meerkat is producing. As you can see, there is a lot going on! 

We are still digging into the data to understand the mechanisms of coordination, but you can check out some of our findings so far here!