SPECTROGRAM BASED CLASSIFICATION OF ANIMAL SOUNDS WITH DEEP LEARNING AND MACHINE LEARNING METHODS

Abstract

The classification of animal sounds is extremely important in fields such as biology, ecology, and environmental science. These classification studies provide great convenience in determining animal populations, monitoring their behavior and defining species. While traditional methods often analyze audio features, deep learning and machine learning techniques have the ability to identify and understand more complex audio features using spectrograms. Deep learning and machine learning can work on spectrograms of audio data. These techniques offer a powerful approach to identifying and classifying animal sounds with more complex characteristics. This approach provides a faster and more automated sorting process than traditional methods, which are often time-consuming and complex. The aim of the study is to develop a model that can automatically classify animal sounds obtained from natural environments without human intervention. In this study, spectrogram images of 577 3-class animal voice data were used. Classification was made with machine learning methods, extracting features with Squeezenet and Inception v3 deep learning methods. The Support Vector Machines (SVM) machine learning model has distinguished animal sounds at a higher rate.