Researchers are cracking the secrets using AI models trained on human speech. Dog Language. the study The study, led by researchers from the University of Michigan, Mexico's National Institute of Astrophysics, and the Institute of Optics and Electronics, was presented at an international conference last week and its promising results suggest that today's AI models could hold the key to understanding, at least to some extent, animal language.
“There's a lot we still don't know about the animals that live in our world,” says Radha Michalcea, director of the University of Michigan's Artificial Intelligence Institute. press release“Advances in AI have the potential to revolutionize our understanding of animal communication, and our findings suggest that we may not need to start from scratch.”
The study uses Wav2Vec2, a state-of-the-art AI speech model, to identify the emotion, gender, and breed behind dog barks. The researchers used two different datasets for training and compared the results: one trained from scratch on dog barks only, and another pre-trained on human vocalizations and then fine-tuned on barks. The model pre-trained on about 1,000 hours of human vocalization recordings performed better. The researchers then fine-tuned the model on a dataset consisting of 74 dog vocalizations (barks): 42 Chihuahuas, 21 French Poodles, and 11 Schnauzers.
Trained on both humans and dogs, the AI model was able to identify dog emotions with 62%, dog breed with 62%, and gender with 69% accuracy, and identify a specific dog within a pack with 50% accuracy. All of these scores outperformed AI models trained on dogs, suggesting that sounds and patterns from human speech could serve as the basis for understanding animals.
In an attempt to uncover the emotions behind dog barks, the researchers hypothesized that dog vocalizations are situational. Existing evidence suggests that the sounds monkeys and prairie dogs make are predictable based on the context of the situation they find themselves in. The emotions the researchers sought to assign to dogs in this study include aggressive barks, normal barks, negative squeaks, and negative growls. Dogs likely experience far more emotions, but these sounds were primarily available in the dataset.
“By using a speech processing model initially trained on human speech, our work opens a new window on how we can leverage what we've built so far in speech processing to begin to understand the nuances of dog barks,” Michalcea said.
The researchers say they hope to test more breeds, emotions, and species in the future to understand the scope of this technology. This is the first time that a human speech model has been used to decipher animal communication, and it could lay the groundwork for understanding animal language. While this study is not conclusive in figuring out the meaning of all dog barks, the researchers see it as a promising step in that direction.
