Human Auditory Representation Learning for cross-dialect bird species recognition

Automated bird species recognition (BSR) is crucial for biodiversity monitoring, but its accuracy is often hampered by geographic variation in bird vocalizations, which is frequently described as dialectal at the population level. This work uses the term dialect region in an operational sense to refer to geographically separated recording regions defined in benchmark corpora that are dominated by different vocal variants across populations, rather than to fine-grained, species-specific dialect types. We aim to address the challenge of generalizing across such dialect-dominated regions by introducing a novel approach leveraging well-established auditory-inspired features known for their computational robustness. We propose Human Auditory Representation Learning (HARL), a framework that integrates Gammatone- and Mel-spectrogram features to capture frequency selectivity and invariance to acoustic variations through their spectral efficiency and empirical success in audio processing. These complementary auditory representations are processed by a dual-stream ResNet50 architecture, with a multi-head attention mechanism to emphasize discriminative spectral–temporal patterns. In cross-dialect evaluation on the D3BV benchmark and cross-site tests on two field datasets (S1 and S2), the approach outperformed strong baselines, raising F1-score by up to 24.31% on D3BV and 28.23% on S1 and S2. Performance remained stable across noise conditions from − 10 to +10 dB signal-to-noise ratio, indicating robustness for real-world deployment. These findings showed that bridging HARL with deep learning delivered a scalable and accurate solution for biodiversity monitoring, enabling reliable species recognition across diverse geographies and acoustic conditions. • Combines human auditory-inspired sound features to enhance bird species recognition. • Uses deep learning with attention to capture complex bird sound patterns. • Improves accuracy in recognizing bird species across regional vocal dialects. • Outperforms baseline methods by over 20% F1-score in cross-dialect settings. • Enables scalable and reliable bird monitoring for ecological conservation.