Frequency Reflects Region in Birdsong Recognition: Quantified by Mutual Information and Mitigated through Adaptive Normalization

Bird vocal dialects, shaped by geographic and environmental variations, present substantial challenges to robust species recognition across regions. To quantify such variability, we apply Maximum Mean Discrepancy (MMD) analysis across multiple birdsong datasets, revealing significant distributional differences between geographically distinct regions. To address these challenges, we propose an adaptive normalization and recognition framework that integrates dynamic feature normalization with a multi-head attention ResNet (MHA-ResNet) classifier. The normalization module adaptively reweights the frequency, time, and channel dimensions based on their task-specific relevance, while the classifier effectively distinguishes both species and regions from the normalized features. This joint design mitigates domain-specific variations while preserving discriminative acoustic cues. To interpret the role of each feature dimension, we use Mutual Information Neural Estimation (MINE) to evaluate their relationships with species and region labels. Experiments conducted on three geographically distinct birdsong datasets show that our method improves species recognition accuracy by an average of 2.9% and region recognition accuracy by 3.0% compared to models without normalization. Mutual information analysis further reveals that frequency features carry the most region-specific information, whereas channel features contribute most to species discrimination. These findings provide new insights into acoustic feature attribution and contribute to biodiversity monitoring across geographic ranges.