Acoustic Classification of Cat Breed Based on Time and Frequency Domain Features

The emerging field of Bioacoustics has been presenting significant research activity lately, and thanks to the use of machine learning methods, several tools and methodologies have been established for identifying certain patterns and meanings in animal vocalizations. Animal sounds can vary over time in intensity and patterns produced between different breeds of the same species, both for physiological reasons and for different emotional states and needs. Pets, such as dogs and cats, are no exception, thus allowing a vocal distinction between breeds. This article studies classification of the cat breed, in particular on the Maine Coon and European Shorthair breed, based on the public audio dataset "CatMeows". To this end, we employed features coming from time and frequency domain capturing relevant information as regard to the present audio structure. Subsequently, audio pattern recognition was carried out by means of k-means clustering, k-NN, and multilayer perceptron learning models. After extensive experiments, we obtained very promising results , with an average accuracy that runs around 98%. In particular, time-domain features presented a strong contribution, as demonstrated by the results using k-means.