Automation of bird voice signal analysis

Objectives. The purpose of the work is to create an experimental software for automated recognition of voice signals, which has the capabilities of long-term round-the-clock and round-the-season monitoring of animal species diversity in selected habitats and ecosystems. Methods. The work uses methods of deep machine learning of convolutional neural networks trained on mel-spectrograms of bird vocalizations, which are built using fast Fourier transform. Results. The process, methods and approaches to training a deep machine learning model for a system of passive acoustic monitoring of bird populations in Belarus are described, as well as the difficulties identified during testing of the software prototype and the results that were achieved. Conclusion. A working prototype of the software for automatic recognition of animal (bird) voice signals is presented. It performs the analysis of acoustic recordings of bird voices with the issue of probabilistic assessment of species belonging to animal vocalizations present in the recordings. The software is aimed at increasing the efficiency of bird monitoring, which ensures the implementation of conservation and research activities based on accurate and up-to-date data on species distribution.