From ‘woof’ to wow: they use AI to decipher dogs’ barking

What kind of information do dog barks contain? That was the main question asked by a group of American and Mexican researchers who created a tool based on artificial intelligence (AI) to better understand what canine vocalizations entail. An AI model developed by Artem Abzaliev, a doctoral student in computer science and engineering at the University of Michigan, in collaboration with scientists at the National Institute of Astrophysics, Optics and Electronics (INAOE) in Puebla, was originally trained on human speech and can be used as a basis to train new systems aimed at animal communication.

The tool, based on a deep learning network and called Wav2vec2, “accepts the audio of a dog’s bark and predicts various properties,” explains Abzaliev, lead author of the study published in The Cornell University Magazine. For its development, a data set of vocalizations of dogs of different breeds, ages and sexes in various contexts captured by Humberto Pérez Espinosa, INAOE researcher, was used. Since 2015, when he worked at the Ensenada Scientific Research and Higher Education Center (CICESE), this computer science expert dedicated two years to capturing barking from more than 100 dogs with the help of veterinary students from the University of Nayarit. “We visited houses where we put the dogs in different situations with the idea of ​​generating positive and negative stimuli to then analyze contrasts through patterns that are generated in vocalizations,” he explains. The purpose of the experiment was to analyze the emotions during these stimuli and then compare them with other different ones and be able to see which one was more similar, “whether a positive reaction or a negative one,” adds Pérez Espinosa.

}Some of the most common tests carried out by his team to record canine responses consisted of a stranger knocking loudly on the door of his home or the dog’s owner pretending to go out for a walk but leaving the house without He, “what commonly causes frustration in dogs, also anguish,” the expert clarifies. To provoke positive stimuli, on the contrary, the pets were given a novel and attractive toy or caresses, provoking a pleasant emotional state in them. All those recordings made over the years by Pérez Espinosa were recorded in a database that his colleague Abzaliev later used to modify a machine learning model, a type of computer algorithm that identifies patterns in large data sets with which the researchers were able to generate representations of acoustic data collected from the dogs and interpret these representations.

The team worked with the Wav2Vec2 model, developed by the company Meta and originally released with human speech data. Wav2Vec2 “uses a method called self-supervised learning, which means it does not require data labeled for human speech,” explains the American researcher. However, since the model was not created to interpret dogs, the team used a more complex method to train it with their samples.

The technique used known as 10-fold cross validation consists of “dividing the data into 10 equal parts. The model is trained using nine of these parts and evaluated on the remaining part; This process is repeated 10 times, alternating the part used for evaluation. In the end, an average of the results obtained in each iteration is calculated,” highlights Pérez Ramírez. As he adds, “the idea is to measure model performance more reliably and reduce the possibility that the result depends on a single split of the data.”

What if artificial intelligence read for us?

Professor Naomi Susan Baron transported us to a world where AI reads and interprets complex texts for us.

According to the results obtained, Wav2Vec2 not only succeeded in four classification tasks (identification of the dog, breed, sex, and situation or context in which it barks) but also outperformed other models trained specifically with data from dog barking, with accuracy figures of up to 70%. Although the vocalizations could be differentiated according to the breed, sex and age of the dog, the easiest thing to identify with the new tool was the bark of each individual, “identifying which dog was barking,” explains Pérez Espinosa. As explained, categories and subcategories are generated in its database. “We know that in the test where a stranger comes to your house, the dogs usually react aggressively.” But between them there are differences in intensity in the reaction, that is, the subclassifications depend on the intensity generated by the dog’s perceived emotion. “So, if we have few categories we have more samples of each category, but if we divide them with greater analysis detail, the data we have decreases. As we apply machine learning techniques, we always depend on having enough information to give examples to the algorithms. By subdividing the samples, the models we train with the data become less precise,” he warns.

Models deep learning to better understand the needs of dogs and save lives

This model is one of the first to use a technique optimized for human speech to help decode animal communication. “There are various techniques in acoustic analysis that are based on some common aspects in the generation of vocalizations, techniques that have worked to analyze human speech have also worked to analyze barking. Even we have tried some developed to analyze music that have given favorable results when analyzing dog vocalization,” explains Pérez Espinosa.

These types of signal processing techniques for the acoustic characterization of human speech serve to describe the properties of the voice in the time or frequency domain. “Many of these techniques used in the characterization of human voice, such as MFCCs, Mel Spectrogram, LPC, F0, are also used to characterize barking, obtaining good results,” the Mexican gives as an example. However, “although other models have been used before to understand barking, models that are not deep learning are usually used,” Abzaliev clarifies. The novelty of the Wav2Vec2 model is that it is a model trained with a lot of human voice data and deep learning (deep neural networks) to represent human voice signals,” Pérez Espinosa clarifies.

In addition to establishing models of human speech as a useful tool to analyze canine and other animal communication, which could substantially help to better understand aspects of biology or ethology, among other fields, this research also has important implications for well-being. animal. Better understanding the nuances of barking could improve how owners or veterinarians interpret and respond to dogs’ emotional and physical needs, thereby improving their care and preventing potentially dangerous situations.