New Dataset Could Transform Study of Mouse Social Communication
Laboratory mice communicate in frequencies too high for humans to hear, creating a persistent challenge for scientists studying their social behavior. CDS PhD student Christopher Ick and colleagues developed a new benchmark dataset that could transform how researchers study rodent vocal interactions, enabling the first large-scale analysis of where sounds originate during social encounters.
The project, a collaboration between researchers at CDS, NYU’s Center for Neural Science, and Princeton’s Neuroscience Institute, aims to understand how rodents interact socially through vocalizations. Previous studies of rodent social behavior have been limited by the need for specialized equipment and invasive procedures that can alter natural behavior patterns. The team’s new approach could enable more naturalistic studies of how rodents communicate in groups.
“No one has really done this sort of dataset for this sort of problem yet,” Ick said. “We’re building methods to essentially generate and annotate more data, which can hopefully tell us more about these things in the future.”
The researchers compiled recordings of over 767,000 rodent vocalizations using several innovative approaches. They attached miniature speakers to robots that moved through testing environments, placed stationary speakers at various locations, and even surgically implanted small earbuds on individual animals to capture vocalizations. This multi-pronged approach allowed them to gather data under different conditions and validate their methods.
Testing focused on two key challenges: determining the precise location of a vocalizing animal, and identifying which animal in a group was making sounds. The researchers evaluated both traditional signal processing methods and newer deep neural networks. While the neural networks achieved sub-centimeter precision in controlled settings with single animals, performance decreased in more complex scenarios with multiple animals interacting. Still, the neural networks consistently outperformed conventional approaches across all conditions.
Ick’s contribution focused on creating accurate acoustic simulations of laboratory environments. By taking detailed measurements of testing chambers and modeling how sound waves bounce within them, he helped generate synthetic data that complemented the real-world recordings. This work built directly on his previous research in sound source localization using simulated data (discussed on the blog previously, here), which demonstrated that models trained on simulated data could perform nearly as well as those trained on real-world data.
The synthetic data proved particularly valuable for understanding how subtle changes in the environment affect sound localization. “Adding a three-degree slant in the walls of the enclosure broadly affects the ability of the model to localize,” Ick said. These insights helped the team refine their methods for different laboratory setups.
The team released their dataset publicly as the Vocal Call Locator (VCL) Benchmark, aiming to spark broader research into rodent social communication. Their work established baseline performance metrics that future researchers can build upon. The benchmark includes not only the recorded data but also tools for generating synthetic data, allowing other researchers to create training data tailored to their specific laboratory environments.
The project emerged when neuroscience researchers noticed Ick’s pioneering work in acoustic simulation in our previous post about his work and invited him to apply his expertise to their study of animal behavior. His methods for generating synthetic training data, originally developed for urban sound monitoring and leak detection on the International Space Station, proved valuable for understanding rodent communication patterns. This cross-disciplinary application of sound localization techniques could open new avenues for studying social behavior in various species.
By Stephen Thomas
