In a pioneering effort, researchers from the University of Michigan, the National Institute of Astrophysics of Mexico, and the Institute of Optics and Electronics of Mexico have employed artificial intelligence models trained in human speech to try to decipher the language of dogs. The results, presented last week at an international conference, suggest that AI models could provide valuable insights into understanding animal languages, albeit in a limited way. "There are so many things we still don't know about the animals that share this world with us," commented Rada Mihalcea, director of the AI Laboratory at the University of Michigan, in a press release. "Advances in AI can revolutionize our understanding of animal communication, and our findings indicate that we don't necessarily have to start from scratch."
Using AI Voice Models for Bark Analysis
The study focuses on using Wav2Vec2, a state-of-the-art AI voice model, to identify the emotion, gender, and breed of dogs based on their barks. To train and validate the model, the researchers used two different datasets. One consisted of a model trained exclusively on dog barks, while the other was a model previously trained on nearly 1000 hours of human speech and later fine-tuned with barks. This approach allowed scientists to compare the effectiveness of both models. The results showed that the AI model including human speech training performed better, successfully identifying a dog's emotion with 62% accuracy, breed with 62% accuracy, and gender with 69% accuracy. Additionally, the model was able to identify a specific dog within a group with 50% accuracy, thus outperforming the model trained solely on dog data.
Context in Vocalization Interpretation
The researchers explored how dog vocalizations might be linked to specific contexts. Previous studies have shown that sounds made by monkeys and prairie dogs can be predicted based on the situation's context. In this study, the aim was to classify canine vocalizations into categories such as aggressive barks, normal barks, whines, and negative growls, to better understand the emotions expressed by the dogs. Although the study does not cover all possible canine emotions, those vocalizations that were most prevalent in the available dataset were selected.
Mihalcea emphasized the significance of using human-trained speech processing models to open new possibilities for understanding dog barks. In the future, the research team plans to expand the study to include a wider variety of breeds, emotions, and species to explore the full scope of this technology. Although the current results are not definitive in fully interpreting dog barks, they represent a promising step toward a greater understanding of animal language.
