Blending Soundscape-Tuned CNNs with a Public Acoustic Baseline for BirdCLEF+ 2026

BirdCLEF+ 2026 evaluates multi-taxon acoustic species identification across 234 classes spanning birds, amphibians, insects, and reptiles, scored by macro-averaged ROC-AUC on a hidden soundscape test set whose public leaderboard is computed on only part of that set during the competition. A single competitor sees a few public-leaderboard readings per day, which raises two practical questions: how to evaluate candidate systems without spending that budget, and how far a public score predicts the private one. We report a single-author entry, built on public resources, that addresses both. The submitted system rank-blends a publicly shared acoustic baseline (a Perch-embedding branch and a five-fold sound-event-detection ensemble) with a weighted triplet of our own soundscape-fine-tuned convolutional networks at weight alpha = 0.10. To rank configurations offline we built a harness on the 66 released labelled soundscape files, with macro-AUC restricted to the 75 species that carry positive labels there; it scores a blend in under a second and replaced the daily leaderboard slot as the primary feedback channel. The entry scored 0.932 macro-AUC on the public leaderboard (1968th of 4092 teams) and 0.917 on the private leaderboard (2415th), and the soundscape-tuned correction helped only at a low blend weight. The two leaderboards did not rank our submissions in the same order, and the harness could not have predicted this: the dominant embedding branch trains on the same labelled soundscapes, so no clean held-out estimate of it exists. We report this generalisation gap and identify a non-leaky cross-validated estimate of the full blend as the change most likely to have closed it.