Performance Evaluation of Bird Sound Classification Models in Realistic Acoustic Environments
Year: 2026
Type: conference-paper
Abstract: Most bird sound classification systems are developed based on clean, single-species recordings. In this work, we assess the robustness of a CNN classifier to overlapping, multispecies acoustic conditions. The baseline single-label CNN yielded 93.4% accuracy for isolated bird calls but its accuracy dropped sharply to 28% when directly tested on mixed-species audio, demonstrating the challenge of species recognition in overlapping ecologies. We propose the use of NMF as a pre-processing step that separates the mixed audio and improves the accuracy to 69%. Further robustness studies using time-stretch augmentation and SpecAugment achieved 65% and 63% accuracy, respectively. The results show that spectral preprocessing and augmentation can significantly improve species recognition performance under acoustic overlap, hence offering a practical pathway towards more effective ecological acoustic monitoring.
Institution: RV University
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
SDG: Life on land
Open Access status: hybrid
OpenAlex ID: W7203608837