Identification of Birds' Voices Using Convolutional Neural Networks Based on Stft and Mel Spectrogram

Abstract

Threats to the climate and global changes in ecological processes remain an urgent problem throughout the world. Therefore, it is important to constantly monitor these changes, in particular, using non-standard approaches. This task can be implemented on the basis of research on bird migration information. One of the effective methods of studying bird migration is the auditory method, which needs improvement. That is why building a model based on machine learning methods that will help to accurately identify the presence of bird voices in an audio file for the purpose of studying bird migrations from a given area is an urgent problem. This paper examines ways of building a machine learning model based on the analysis of spectrograms, which will help to accurately identify the presence of bird voices in an audio file for the purpose of studying the migration of birds in a certain area. The research involves the collection and analysis of audio files that can be used to identify characteristics that will identify the sound of the files as birdsong or the absence of sound in the file. The use of the CNN model for the classification of the presence of bird voices in an audio file is demonstrated. Special attention is paid to the effectiveness and accuracy of the CNN model in the classification of sounds in audio files, which allows you to compare and choose the best classifier for a given type of file and model. Analysis of the effectiveness and accuracy of the CNN model in the classification of sounds in audio files showed that the use of Mel-spectrograms is better than the use of STFT-spectrograms for studying the classification of the presence of bird sounds in the environment. The classification accuracy of the model trained on the basis of Mel spectrograms was 72 %, which is 8 % better than the accuracy of the model trained on STFT spectrograms.