Woof! Computer Speaks Dog

Feb. 4, 2008 -- Dog owners may think they can recognize their pet's bark. But a new study shows that a computer is far better at identifying a canine's woofs, arfs, and ruffs than a human.

The computer, which relies on artificial intelligence and machine learning methods, could provide a new tool for automatically analyzing large amounts of data typically recorded during behavioral studies.

"Animals communicate very different emotional states by very different vocal sequences," said Csaba Molnár, an ethologist from Eötvös Loránd University in Budapest, Hungary.

"You could use this method to analyze other vocal signals, such as bird songs, or categorize video recordings. You could even use it to categorize human speech and dialogues."

The software, which was developed by collaborators at the Sony Computer Science Laboratory in Paris, France, analyzed more than 6,000 barks from 14 Hungarian sheepdogs known as Mudis.

The researchers recorded barks in six different behavioral categories: "stranger" (where an unknown person appeared at the house in the absence of the owner); "fight" (where a trainer encourages the dog to bark aggressively); "walk" (where the owner prepares for a walk outside); "alone" (where the owner tied the dog to a tree and walked out of sight); "ball" (where the owner held a favorite toy); and "play" (where the owner played a familiar game with the dog).

After the barks were collected, they were transferred to a computer and digitized. Artificial intelligence software is then used in a two-stage process to first learn the more than 100 acoustic features of the different kinds of barks and then use that knowledge to code, classify and evaluate them.

In one experiment, the software correctly classified the barks in 43 percent of cases. "Fight" and "stranger" barks were the easiest to recognize, while "play" barks were more difficult. When matched against a human's ability to do the same, the computer's success was about the same.