A Novel Deep Learning Approach for Classification of Bird Sound Using Mel Frequency Cepstral Coefficients

Abstract

 Monitoring animal populations is one important matter to better understand changes in their
population, behavior, and biodiversity. Bird sounds are the main tool to classify bird species acoustically. The
sounds of birds are an indicator for ecologists as it responds to changes in their environment. The recognition
among a variety of bird species to get important features is computationally expensive. With the unbalanced
classes and scarcity of training data, the performance accuracy is degrading. This paper aims to classify species of
birds using lightweight convolutional neural networks (LWCNNs) basis on using a spectrogram image of Brazilian
bird sounds as a dataset. For extracting spectrogram images, Mel Frequency Cepstral Coefficient (MFCC)
algorithm is used. To prove the high performance of the classifier, ten species of birds with 10,000 spectrogram
images are provided to the classifier. Our LWCNN model achieved a training and testing accuracy of 99.68 % and
92.80 % respectively in 10.54 min with 5 epochs.