Researchers are using AI models trained on human speech to decode the secret language of dogs. The study comes from researchers at the University of Michigan, Mexico’s National Institute of Astrophysics, and the Optics and Electronics Institute. The promising results, presented last week at an international conference, show how today’s AI models could be a key to understanding animal languages, at least to some degree.
“There is so much we don’t yet know about the animals that share this world with us,” said Rada Mihalcea, the director of the University of Michigan’s AI Laboratory in a press release. “Advances in AI can be used to revolutionize our understanding of animal communication, and our findings suggest that we may not have to start from scratch.”
The study utilizes a state-of-the-art AI speech model, Wav2Vec2, to identify the emotion, gender, and breed of a dog behind any given bark. Researchers two different data sets for training and compared the results: one was trained from scratch on just dog barks, and one was pre-trained on human speech and then fine-tuned on barks. The model pre-trained on nearly 1,000 hours of human speech recordings did better. Researchers then fine-tuned that model on a data set consisting of vocalizations (barks) from 74 dogs: 42 Chihuahua, 21 French Poodles, and 11 Schnauzer.
This AI model trained on humans and dogs was able to identify a dog’s emotion with 62% accuracy, breed with 62% accuracy, gender with 69% accuracy, and identify a particular dog out of a bunch with 50% accuracy. All of these scores outpaced the AI model just trained on dogs, which suggests that sound and patterns derived from human speech can potentially serve as a foundation for understanding animals.
In trying to unpack the emotion behind a dog bark, the researchers hypothesize that a dog’s vocalization is related to its context. Existing evidence suggests the sounds that monkeys and prairie dogs make can be predicted based on the context of the situation they’re in. Some of the emotions researchers try to assign to dogs in this study are aggressive barking, normal barking, negative squeals, and negative grunts. While dogs likely experience vastly more emotions, these noises were largely available in their dataset.
“By using speech processing models initially trained on human speech, our research opens a new window into how we can leverage what we built so far in speech processing to start understanding the nuances of dog barks,” said Mihalcea.
Moving forward, the researchers say they would like to test more breeds, emotions, and species to understand the extent of this technology. This is the first time human speech models have been used to decode animal communications, and it could lay the groundwork for understanding animal language. While this study is certainly not definitive in unpacking the meaning of all dog barks, researchers see it as a promising step in that direction.
