🐦 BirdCLEF 2026
Acoustic species identification in the Pantanal wetlands. 234 species from 60-second soundscape recordings.
Quickstart
uv sync # install
uv run python scripts/train.py --data-dir data/competition # train
uv run python scripts/evaluate.py --data-dir data/competition # evaluate
uv run python scripts/export_notebook.py # export notebook
uv run pytest # tests
All scripts default to configs/base.yaml. Pass --config to use a different config.
Architecture
The pipeline uses Perch v2 embeddings (Google's pre-trained bird vocalization classifier) fed into a ProtoSSM v4 (Prototypical State Space Model) with:
- Mamba-style selective SSM for temporal modeling
- Cross-attention for non-local patterns
- Per-class prototypical learning
- Gated fusion with Perch logits
- MLP probe ensemble for per-class refinement
- Site/hour prior probability tables
Competition Constraints
- CPU-only, 90 minute wall time
- No internet during submission
- Self-contained notebook required