Improving animal welfare by means of automatic monitoring of goat vocalizations: development of a bleating recognition filter

Ensuring high levels of animal welfare is fundamental for a better and more sustainable use of resources in animal farming. In fact, high levels of animal welfare are related to higher immune response, lower risk of disease, lower use of drugs and antibiotics, and, therefore, fewer drug residues in the environment, thus leading to better global health. The VOCAPRA project aims to study goats’ vocalizations in order to improve human-animal communication and provide farmers with an early warning tool to detect poor animal welfare conditions and intervene to meet the animals’ needs. To this aim, the development of a filter, i.e. a non-linear supervised algorithm, was initially required for distinguishing goat vocalizations from all other sounds. For a whole year starting from April 2021, we recorded sounds in four goat farms, using 18 acoustic sensors. We selected 150 audio tracks and classified them as: noise (n=60), adult goats’ vocalizations (n=50), and kids’ vocalizations (n=40), based on the prevailing sound. Then, the tracks classed as adult goats’ vocalizations were edited to isolate only bleats. Using a sliding window technique, each track was then split into smaller tracks, whose acoustic characteristics were extracted and subjected to a random forest classifier algorithm, obtained from the sklearn.ensemble library, to quantify the number of bleats. On the basis of the number of bleats we then assigned a score from 0 (no bleats) to 1 (only bleats) to each smaller track. The algorithm was subjected to k-validation training (k=5) using for each k the 80% of all classified audio tracks, while the remaining 20% was used to test the algorithm. We reached an algorithm accuracy of 82%, suggesting the possibility of using it to select goats’ vocalizations’, which is the first step for developing an IT tool for monitoring goat welfare based on vocalizations.