Researchers aim to translate dog sounds with AI

Ever wonder what your dog is trying to say? Well, a University of Texas at Arlington researcher is aiming to turn barks, howls and whimpers of man’s best friend into intelligible speech — a kind of Rosetta Stone of woof.

Computer scientist Kenny Zhu has built what he says is the world’s largest video and audio catalog of canine vocalisations. In papers published this year, Zhu and his colleagues at the university report potential phonemes — the smallest units of sound — and word-like patterns that could one day be turned into full sentences understandable to humans.

“The ultimate goal is to make a translator where you can talk freely with your pet,” said Zhu, a professor of computer science and engineering at UT Arlington. “We can already do instantaneous communication between human languages. Perhaps in the future we can do the same with animals.”

AI interprets dog: Humans have long wanted to talk to animals, and in the last century, scientists have tried: from teaching great apes sign language and English to bottlenose dolphins.

Zhu’s fascination with animal communication began in Nanjing, China, where he spent his childhood surrounded by dogs, ducks, chickens and the occasional hedgehog. He often wondered what the animals were saying to each other, though his curiosity cooled over time.

It wasn’t until decades later, when he was watching a BBC documentary on whale and dolphin communication, that questions from his childhood reemerged. The documentary showed how long and hard it was to record and decode whale and dolphin exchanges. But with artificial intelligence, Zhu thought, there had to be an easier way to translate animal speech. With his background in natural language processing and AI development, he felt up to the task.

For his first project, Zhu wanted to see if a language model could hear a difference between Shiba Inus in Japan and in the United States. He and his colleagues mined dog videos posted on YouTube for the test. After it didn’t reveal any doggy dialect split, Zhu and his colleagues compiled hundreds of hours of synced audio and video to train an AI model to separate canine vocalisations into discrete phonemes.

Deciphering the vocalisations involves both sound and context, as a dog’s bark or whine may be tied to its situation, Zhu said. If a term aligns with the dog’s activity, that correlation signals potential meaning.

So far, the researchers have transcribed about 50 hours of barks into syllables. They have identified some possible words, like cat, cage and leash, and how these words seem to sound different based on the dog breed. They have also identified how a dog’s linguistic capability appears to change as it ages. In one study, Zhu and his colleagues found that as a husky grows older, its bark lasts longer and potentially becomes more sophisticated.

Dr. Doolittle at your fingertips: This effort isn’t just about chatting with Fido like your next-door neighbor: It could also help flag early clues about your dog’s health, Zhu said. If a dog experiences any mental or physical changes, a smartphone app or other device outfitted with a dog translator could inform the owner.

To a similar end, Zhu is working on decoding cats. He’s drafting a proposal to the Morris Animal Foundation for a study investigating whether a cat’s vocalisations can provide insight into its mental state or behaviour. Another one of Zhu’s projects, with Texas A&M University, is tackling the sounds of cattle. Dozens of cows in monitored pens at the university have been recorded 24/7 for audio and video for over two months.

The data will be compared to the animals’ veterinary records to see how it correlates with their health. Zhu and his collaborators hypothesize that herd small talk may carry cues about bovine well-being.

By analyzing those vocal patterns for linguistic structure, they hope to spot illness before a human sees a sick cow.

They aren’t the only ones using AI to decipher animal speech. At the University of Michigan, researchers have processed dog barks using AI models originally trained on human speech, and at Virginia Tech, scientists are building an AI system to decode cow vocalisations. Meanwhile, a cottage industry of AI-powered dog collars and “cat translator” apps has sprung up, promising users the ability to better understand the needs of their pets.