News

Can AI tell you what your dog is saying? U-M researchers are testing it

Posted
and last updated

ANN ARBOR, Mich. (WXYZ) — Dog lovers, imagine a puppy that barks, growls or groans at you but you know exactly what they're trying to tell you.

While we're not there yet, the University of Michigan is taking baby steps into understanding animal vocalizations, and they're using artificial intelligence to understand human speech to do it.

You may remember the old TV show "Lassie." All it took was a couple of barks, and Lassie's owners knew little Timmy was stuck in a well somewhere and needed rescuing.

That's the direction the U of M researchers are going, and dog owners say they love anything that gives a closer look into what's going on behind those puppy dog eyes.

Christian Hauser and his dog, Lottie, are two buddies we caught up with to ask, what if AI could help you understand your dog's barks.

Screenshot 2024-07-02 at 2.33.18 PM.png

"I always say if I could have one thing, I would want to be able to hear her talk for one day, to understand and maybe see what she has to say," Hauser said.

Hauser isn't a lone wolf.

"I need it. I would love it. I would love it. I think it would be great," Christine McPherson said. "Anything to better understand them."

U-M doctoral student and researcher Artem Abzaliev, along with Rada Mihalcea, a professor of computer science and engineering, are working toward that.

The pair used an AI model called Wav2Vec2. It was created to interpret human speech for things like voice-to-text and language translation. They then applied that model to Hundreds Vocalizations Research Partners in Mexico recorded from 74 dogs.

"They would make the owner play with the dog and while the dog was playing, they would record every sound the dog makes," Abzaliev said. "From the computer science perspective, similar way we treat human speech or any sound, animal vocalizations is just sound, so any other methods developed for the audio process can be applied to it. We didn’t expect it to work so it was very surprising to us that it worked."

Screenshot 2024-07-03 at 6.06.25 AM.png
Artem Abzaliev, PHD student at U of M

They set out to see if the AI model could determine the dog's sex and breed, if a bar can be identified, and the context of the vocalization. Whether it was playoff or if it was more protective.

Sex was the only measure the AI model couldn't detect from barks.

While Abzaliev said the bond between owners and dogs is strong enough to know what they're saying, animal rescuers could benefit.

"In a shelter for instance, not every dog has the opportunity to establish those deep connections, so maybe we can predict anxiety in dogs in the shelter," Abzaliev said.

This is just the start of the ground-barking research. Abzaliev hopes to add a video element to further understand what our dogs are saying to us.