Artificial intelligence: tell me what my dog wants to tell me
Researchers have applied an AI model for speech processing to dogs barking. The result: the artificial intelligence even detected the animals' mood.
What does a dog's barking say about its current mood? An AI trained on human voices was able to answer this question.
(Image: Zivica Kerkez/Shutterstock.com)
Does the dog just want to play or is it aggressive? An artificial intelligence (AI) was able to answer this question, a research team at the University of Michigan has discovered. They applied a language model that is actually trained on human language to dog barking and discovered that it works: It works. The tool recognizes whether the dog's barking is playful or aggressive.
The team fed the model with 74 sounds from dogs of different breeds, ages and genders, which were recorded in different situations. The recordings were made in controlled situations. For example, barking was recorded when the doorbell rang several times, when the dog was playing or when an attack on the dog owners was simulated.
AI recognizes breed, age, mood and gender
The same models were also able to read out how old the dog is, what breed it belongs to and what gender it is from the animal sounds. With the Wav2Vec2 model used by the team, they were able to interpret the collected data. The model fulfilled four classification tests (voice recognition, breed, gender and mood) and also outperformed other models that had been specially trained with dog barking with an accuracy of up to 70 percent.
By using a model that was actually designed for human speech, the researchers wanted to solve a problem with the training data. In contrast to human voice recordings, there is significantly less material of animal communication. "Animal sounds are much more difficult to capture and record," said Artem Abzaliev, lead author and PhD student in computer science and engineering at the University of Michigan, according to a statement from the university. Either the recording would have to be done passively in the wild or with owner permission in the case of pets.
AI model was trained to interpret vocal subtleties
The research team therefore used the models on which current voice-controlled technologies are based, including translation services. These are already designed to perceive and interpret subtleties such as voice, tone of voice, pitch or accents. "These models can learn and decode the incredibly complex patterns of human speech," said Abzaliev. "We wanted to know whether we could also use this ability to recognize and interpret dog barking."
The team sees the findings as valuable not only for the use of language models of human communication for use in analyzing animal communication. They also provide important insights for animal welfare, as a better understanding can also lead to better treatment of animals - not only an achievement for animal welfare, but also important for defusing dangerous situations.
The team presented the results at the Joint International Conference on Computational Linguistics, Language Resources and Evaluation.
(are)
