Belarusian bird acoustic recognition: data preparation and model training process

The issue of substantial labor and time demands for monitoring bird species diversity and range changes, especially in developing countries, invites novel technological solutions. The recent advancements in machine learning (ML) have led to breakthroughs in AI-based data processing, including tools for automated passive acoustic monitoring (PAM) that utilize on-site bird vocalizations. Here we describe our preliminary results and difficulties encountered when developing an EfficientNetB3-based model for a PAM system to monitor bird diversity in the forested areas of interest in Belarus. A novel dataset of bird vocalizations from Eastern Europe, processed and converted into mel-spectrograms allowed us to achieve a respectable f1-scores (>0.9) in tests for certain species such as nightjar and nutcracker. However, the overall score (0.52) for the 116 species of interest was unacceptably low. Further testing with a more specialized dataset allowed us to determine that the problem lies with the peculiarities of species, and is not limited to species with complex vocalizations. We hypothesize that model overfitting to specific vocalization signals may be one of the main causes. Additionally, certain species require a thorough coverage of their vocalization diversity in the dataset.