“Who is calling?” A novel algorithm to identify individual marmosets based on their call
Who’s calling? Marmosets are highly social and vocal monkeys, making them ideal models for studying vocal communication, and eventually its evolution. But this also comes with its share of problems, as analyzing these complex communication signals can be tricky. Fortunately, a team of NCCR researchers headed by Prof. Judith M. Burkart has found a solution to this.
Marmoset monkeys © Judith M. Burkart
There is a complexity to marmoset communication: the monkeys are very voluble –, they produce a lot of complex calls –, and they live in groups –, which means many individuals produce vocalizations at the same time. “Because marmosets live in noisy groups, it makes it hard to determine which individual in the group emitted a specific call,” explains Nikhil Phaniraj, PhD candidate in Burkart’s team and first author of a recently published study tackling this problem. “Knowing which individual vocalized would be very valuable for research.”
Identity crisis in marmosets
Burkart and her team study the vocal communication of marmoset monkeys to understand the origins and the evolution of language. “For this, we have to study how vocalizations develop in individual marmosets, whether and how these vocalizations change with time, and how individuals respond to vocalizations from other monkeys,” says Phaniraj. Previously, to achieve this, researchers would isolate an individual from the group and study its vocalizations. “But this makes it unnatural because when isolated, the marmosets do not fully display their communication skills,” Phaniraj deplores. It is therefore key for the researchers to study the monkeys in their social group – which would require a method allowing them to differentiate them when they vocalize in groups.
We have your caller ID!
The researchers grasped the nettle and developed a novel machine-learning algorithm that can determine the identity of the monkey from its vocalization alone. The pipeline first performs the easier task of determining the sex of the monkey from patterns in the vocalization, and later determines the individual, thanks to the extraction of more than 7,700 features contained in each call. “This enables it to attain superior accuracy compared to traditional approaches aiming to directly determine the individual,” comments Phaniraj. The pipeline was tested on a set of 20 individuals (10 males and 10 females) and showed very promising results.
Like with any machine-learning algorithm, the most important thing to obtain accurate results is the quality of the dataset on which the algorithm will learn. “A large part of the project was focused on coming up with good datasets to train the machine-learning algorithm on,” says Phaniraj. “We wanted to make sure our data set was balanced – meaning calls of all individuals and all sexes are represented equally.” Now that their pipeline is up and running, the technology can be applied to the lab’s projects. For instance, the researchers already use it to determine the ‘prosociality’ of individual monkeys. “One way marmosets help other members of the group (i.e., being ‘prosocial’) is by producing food calls when they find food so that they can inform other members and share it with them,” explains Burkart. “By studying which members of the group do and don’t give food calls when they find food, we can estimate how prosocial they individually are.”
In the future, the same pipeline could be extended to differentiate individuals of other group-living species.
Phaniraj Nikhil, Wierucka Kaja, Zürcher Yvonne and Burkart Judith M. 2023. Who is calling? Optimizing source identification from marmoset vocalizations with hierarchical machine learning classifiers. J. R. Soc. Interface. 20: 20230399. http://doi.org/10.1098/rsif.2023.0399.