The Evolutionary Structure of Acoustic Learnability: A Deep Learning Approach to Neotropical Birdsong

Passive Acoustic Monitoring offers a scalable solution for biodiversity assessment in the Neotropics, but classifying hundreds of sympatric species from complex soundscapes remains a major challenge. Here, we develop a deep learning framework for large-scale avian sound classification, training convolutional neural networks on recordings from 667 Neotropical bird species across northern South America. Our ResNet-based architecture achieves high performance (86.09% accuracy; 84.97% F1-score), establishing a new regional benchmark. Beyond prediction, we apply explainability and uncertainty quantification to evaluate how model confidence aligns with acoustic structure and biologically meaningful ambiguity. To explore the biological drivers of classification performance, we test three macroevolutionary hypotheses--morphological constraint, acoustic adaptation, and cultural evolution--by linking model outcomes with species-level traits under phylogenetic control. We find that neither morphology nor habitat structure predicts accuracy once evolutionary history is accounted for. Instead, geographic range size shows a consistent negative relationship with performance, suggesting a role for intraspecific vocal diversity in shaping acoustic distinctiveness. Overall, our results demonstrate that deep learning can serve not only as a tool for biodiversity monitoring, but also as a framework to test evolutionary hypotheses and interpret the structure of animal communication through biologically meaningful model behavior.