BirdCLEF+ 2026 — soundscape-tuned model weights
ONNX weights for a single-author entry to BirdCLEF+ 2026 (LifeCLEF, CLEF 2026). Code and the CLEF working note: https://github.com/gcol33/bird-clef-2026.
The submitted system blends a publicly shared acoustic baseline (a Perch-embedding probe branch and a five-fold SED ensemble, not included here) with the weighted triplet of soundscape-fine-tuned convolutional networks released below, rank-blended at weight 0.10. Final scores: public 0.93215 (1968 / 4092), private 0.91710 (2415 / 4092), macro-averaged ROC-AUC over the 234 classes.
Files
| File | Backbone | Role | Triplet weight |
|---|---|---|---|
b3_ss.onnx |
EfficientNet-B3 | soundscape fine-tuned | 0.45 |
b3_pseudo_ss.onnx |
EfficientNet-B3 | soundscape fine-tuned + pseudo-labels | 0.40 |
convnext_small_ss.onnx |
ConvNeXt-Small | soundscape fine-tuned | 0.15 |
crossyear_b3/model.onnx (+ .data) |
EfficientNet-B3 | cross-year pretrain then fine-tune (development track) | — |
label_order.json |
— | the 234 class identifiers in output-column order | — |
crossyear_b3/model.onnx uses ONNX external-data; keep model.onnx.data alongside it.
Input
32 kHz mono audio, 5-second windows, 128-band log-mel spectrogram (n_fft = 1024, hop = 320,
50–14000 Hz, power, peak-normalised per window), single channel. Output: 234 per-class scores in the
order given by label_order.json.
License
Weights released under CC-BY-4.0, subject to upstream dataset terms. The competition data is governed by the Kaggle competition rules and is not redistributed here.