Reliability-aware regional calibration for fine-grained bird sound recognition with a frozen BirdNET backbone

Abstract Background Passive acoustic monitoring produces large volumes of bird sound recordings, but automated species recognition for a regional checklist must handle sparsely represented species and provide a reliability score for each prediction, so that limited expert review effort can be allocated to uncertain cases. Results We present RRC-BirdNET, a regional calibration framework built on a frozen BirdNET backbone. A lightweight window-score head projects acoustic embeddings into regional species scores and aggregates them through per-class-temperature log-mean-exp pooling (LME-PC); balanced-softmax training incorporates class-frequency information; and a confidence-based reliability gate provides a development-selected operating point for automatic reporting versus expert review. On 3929 provenance-audited recordings covering 276 species, the five-head ensemble achieves 84.05% top-1 and 79.94% balanced accuracy on 815 test recordings; individual heads achieve 83.51±0.34% top-1 across five seeds. On the 787-recording subset with direct BirdNET taxon mappings, the regional ensemble achieves 84.24% top-1, compared with 64.68% for checklist-restricted frozen scores. The reliability gate reaches an error-detection AUROC of 0.921; at a development target of 5% accepted-set error it retains 80.7% of test predictions with 95.6% observed accuracy. One head occupies 1.14 MB and its forward pass takes 0.54 ms per 96-window input on a CPU workstation. Conclusions Compact regional adaptation of a shared acoustic representation supports species-level monitoring with selective reporting: the checklist readout improves over frozen global scores, the long-tail correction improves few-shot recall, and the reliability gate makes the trade-off between automation and expert review explicit and measurable. The framework runs locally without network connectivity, supporting deployment in field monitoring workflows. sons. Species-level information is valuable for biodiversity assessment (IPBES, 2019) and for the management of wetlands along the East Asian–Australasian Flyway (Studds et al., 2017). Advances in acoustic recorders and deep learning enable these observations to be collected and analysed at increasing scale (Stowell, 2022).