Animal communication: Using new AI methods to decipher the language of mice

Analyzing animal vocalizations is an extremely challenging task in behavioral biology—and especially for studying the ultrasonic calls of house mice (Mus musculus). Bioacoustic analyses are challenging for many species due to the large volumes of data whose manual evaluation has, until now, required considerable time and labor. Current approaches from bioacoustics, artificial intelligence (AI), and machine learning (ML) offer opportunities to automate these processes and increase the explanatory power of analyses. “For a long time, research on animal communication was constrained by the effort involved in preparing and analyzing audio data. Today, new AI-based methods enable far more efficient and precise analyses,” explains study co-last author Sarah M. Zala of KLIVV.

More error culture in data processing and analysis

In their research based on existing studies, the scientists provide a comprehensive overview of currently available tools for processing and analyzing mouse vocalizations. The research team paid particular attention to the individual steps of data processing—from recording and signal preprocessing to the automatic detection of ultrasonic vocalizations and their classification and analysis. For future studies, Zala says, the goal is to “minimize errors at each processing step and prevent their propagation along the analysis pipeline.”

Opportunities and limits of automation

A comparison of established signal-processing techniques with machine learning methods shows that automated detection of vocalizations has made considerable progress in recent years. At the same time, challenges remain-for example, in suppressing background noise, reliably tracking frequency contours, and extracting biologically relevant features.

Reference datasets: Human as the gold standard

Another focus of the study was the classification of mouse vocalizations. This is especially complex because some call types are clearly distinct, while others show gradual transitions. The researchers therefore emphasize the importance of high-quality, manually curated datasets as references for automated methods. “For this reason, we propose a new hierarchical framework for classifying ultrasonic vocalizations,” emphasizes study co-last author Dustin J. Penn of KLIVV, adding: “Automated methods are only as good as the data on which they are trained and validated. Carefully constructed reference datasets are, and will remain, the indispensable gold standard.”

High potential for new insights into animal communication and social behavior

The researchers also note that open data collections and standardized analytical procedures will play a central role in future scientific progress. In the long term, AI-based methods could help decode the information contained in animal signals and thereby yield new insights into animal communication and social behavior.

Even now, however, the study’s findings are an important contribution to behavioral research, says Penn: “Our results provide a clear and practical guide for researchers in behavioral biology and bioacoustics and pave the way for the next generation of automated tools for analyzing animal communication.”

 

The article „Bioacoustic processing and analyses of mouse vocalizations: Current methods and future directions“ by Reyhaneh Abbasi, Doris Nicolakis, Maria Adelaide Marconi, Teresa Klaus, Bettina Wernisch, Maja Peng, Peter Balazs, Dustin J. Penn and Sarah M. Zala was published in „Behavioural Brain Research“.

Scientific article


Scientific contact:
Sarah Zala PhD.
Konrad-Lorenz-Institut für Vergleichende Verhaltensforschung (KLIVV)
Veterinärmedizinische Universität Wien (Vetmeduni)
Sarah.Zala@vetmeduni.ac.at