- June 7, 2024
- AI in Business
Have you ever wanted to know what your dog needs to communicate? Researchers at the University of Michigan are exploring the potential of artificial intelligence (AI) in creating a system that can distinguish whether a dog’s barking means playfulness or aggression.
These AI models can also extract additional information from animal sounds, determining the age, breed, and gender of the animal. In collaboration with Mexico’s National Institute of Astrophysics, Optics, and Electronics (INAOE in Puebla), the study shows that artificial intelligence systems originally trained on human language can be adapted to decipher animal communication. The findings were presented at the Joint International Conference on Computational Linguistics, Language Resources, and Assessment and published on the arXiv preprints server.
“By using speech processing models trained on human language, our research provides new insights into the use of these tools to understand the subtleties of dog barking,” said Rada Michalcha, the Janice M. Jenkins Collegiate Professor of Computer Science and Engineering and director of U-M AI. Laboratory. “We still have a lot to learn about the animals we live with on our planet. The development of artificial intelligence can greatly improve our understanding of animal communication, and our results suggest that we may not have to start from scratch.”
A major obstacle to the development of AI models to analyze animal sounds is the scarcity of publicly available data. Although there are many resources for human language, collecting data from animals is much more difficult. “Requiring and recording animal sounds is more difficult from a logistical standpoint,” noted Artem Abzaliev, lead author and a U-M computer science and engineering doctoral student. “They must be passively recorded in their natural habitat or, for pets, with the owner’s consent.”
Due to data limitations, creating methods for analyzing dog vocalizations has been a challenging task, and existing models are limited by a lack of training data. The researchers solved these problems by adapting an existing model originally developed to analyze human speech. This strategy has allowed researchers to use reliable models that form the basis of a variety of voice technologies, such as voice-to-text and speech translation. These models are trained to detect nuances of human speech, including tone, pitch, and accent, and convert that information into a format that computers can recognize.
“These models are designed to learn and encode complex patterns of human language and speech,” Abzaliev said. “We wanted to determine whether this ability could be applied to the interpretation of dog barking.” The team used a dataset of vocalizations from 74 dogs of different breeds, ages, and sexes recorded in different contexts. INAOE’s Humberto Pérez-Espinosa led the team that collected the dataset. Abzaliev then modified a machine-learning model that identifies patterns in large data sets using these records. They chose a speech model known as Wav2Vec2, originally trained on human speech data.
Using this model, researchers could generate representations of the dogs’ acoustic data and interpret those representations. They found that Wav2Vec2 not only coped with four classification tasks but also outperformed other models trained specifically for dog vocalizations, achieving an accuracy level of up to 70%. “This is an innovation in the sense that we have adapted methods optimized for human language to decode animal communication,” Michalca said. “Our results show that human speech patterns can provide a basis for analyzing the acoustic patterns of other sounds, such as animal vocalizations.”
In addition to establishing that human speech patterns are valuable for analyzing animal communication—which could benefit biologists, animal behaviorists, and the like—this research has significant implications for animal welfare. Understanding the nuances of dog barking could improve how people interpret and respond to the emotional and physical needs of dogs, improving their care and potentially preventing dangerous situations, researchers concluded.
