Development of a Bird Song Identification Model Using a Convolutional Neural Network (CNN)

Abstract

Birds have a very large number of species that show remarkable diversity, but this diversity is threatened due to human activities and natural phenomena. Identifying bird species that still exist in an ecosystem is an initial action that can be taken to protect this bird diversity. In this study, a Convolutional Neural Network (CNN) model was built to automate the process of identifying bird species in bird song data. In the experiments conducted, 24 CNN models were built, which had differences in architectures, feature extraction techniques, and types of input data. There are three architectures used: AlexNet, DenseNet, and VGG. There are four feature extraction techniques used: mel-spectrogram, harmonic component-based mel-spectrogram, percussive component-based mel-spectrogram, and MFCC. As for the types of data used, there are two: data that has been cleaned (clean data) and data that has not been cleaned (raw data). Based on the experiment's results, the CNN model with the best performance is obtained with the DenseNet architecture, feature extraction with percussive component-based mel-spectrogram, and clean data as model input. This evaluation process is carried out using two types of testing methods: testing using chunk data and testing with complete data. The accuracy of the best model reaches 85.67% for testing with chunk data (method 1) and 90.42% for testing with complete data (method 2).