Tricia Jonina B. Alcisto, William Caleb Arcena Perez, Angie M. Ceniza-Canillo
Bird sounds are a rich source of information that can significantly enhance our understanding of avian biodiversity, behavior, and their ecological roles. Ornithological research is driven by the need to understand avian biodiversity and their ecological dynamics. It relies heavily on accurate identifications and analysis of bird sounds. The manual analysis of bird songs is time-consuming and is prone to human errors, limiting the scope and efficiency of research efforts. Despite these problems, research efforts and technological advancement has made it possible to automate the process of species classification based on their sound. This study presents a novel approach for automated bird sound classification using Mel-Frequency Cepstral Coefficients combined with Spectral Contrast, creating a two-channel MTF image representation processed through a CNN model. This approach captures both timbral and harmonic characteristics in a unified image representation. Ten bird species were studied and will be classified: (1) barn swallow, (2) Eurasian coot, (3) Gray Heron, (4) Gray-Headed Canary Flycatcher, (5) Kentish Plover, (6) Large-Billed Crow, (7) Purple Heron, (8) Rock Pigeon, (9) Whiskered Tern, and (10) Wood Sandpiper. Each species represents a mix of habitats. Each gathered from BirdCLEF 2024 data set, a large-scale benchmark for bird sound classification. The model's performance resulted in an 89.9% score in the accuracy metric, 89.5% in precision, 88.9% in recall, and lastly 89.2% for the F1-score. The approach outperformed its ablation study with a performance gain of approximately 4%, and 3.6% over the baseline approach. © 2025 IEEE.
University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu, Philippines
