[El1as42]

Summary

• Model menu in the Demo tab: Just-Bird stays the default; All-In-One (282 classes) can be picked from the menu at the top. With All-In-One the table names the species of each box.
• Smaller ONNX graphs: onnx_export.py no longer stores the 15.7 MB STFT kernel; the graph rebuilds it when the model loads, bit-identical. Just-Bird_fp16.onnx shrinks from 35 MB to 20 MB, the new All-In-One_fp16.onnx is 51 MB instead of 66 MB.
• No species badge: the species-282 badge is gone from the README and the docs landing page. The recommended workflow combines BirdBox boxes with BirdNET for labels, so BirdBox's species count should not be up front. The landing page's Python badge now links to 3.12.13 like the README.

Model menu

• The menu lists docs/models/manifest.json; the first entry is the default. Switching runs detection again with the model's suggested slider values. Drawn boxes stay; verified, moved or removed detected boxes are reset, with a notice.
• All-In-One merges by song reconstruction, not window fusion. songs.js ports reconstruct_songs_tensor, so the sliders stay instant; the second slider becomes Song gap (default 0.1 s).
• With more than one class, a Species column (common name and eBird code) and the box tooltip name the species. The manifest gains song_gap and common_names (from src/config.py).
• The JSON export records the model's merge method and gap (phrase_gap for fusion, song_gap_threshold for reconstruct, as detect_birds.py writes them).
• The page title reads "BirdBox in your browser" now that the model can change.

Rebuilding the STFT kernel at load

• The kernel's phase at bin k and sample n only depends on (k * n) % 2048. The file stores one period of cosine and sine values plus the window (about 50 KB); dft_kernel_builder adds nodes that do the lookup and products in float64 and cast to float32, as dft_convolution_kernel does. The loop body reads the kernel as an outer-scope value.
• The export checks in numpy that the lookup reproduces the kernel exactly and stops otherwise.
• Every export with onnx_export.py uses it, for every model and precision. fp32 graphs shrink too (Just-Bird 53 → 38 MB, All-In-One 114 → 99 MB).

Results

Graph (13 recordings, tests/onnx_graph_parity.py)	Boxes of detect_birds.py with the .pt model matched	Mean IoU
Just-Bird_fp16.onnx	99.8%	0.995
All-In-One_fp16.onnx	99.7%, same species (1066 of 1069)	0.997
All-In-One_fp32.onnx (reference)	100%	1.000

Test plan

• Old and new graphs give bit-identical detections and raw_all on tests/test.wav, in ONNX Runtime (Python) and ONNX Runtime Web 1.20.1 (WASM); same load time
• tests/onnx_graph_parity.py for both demo models (table above)
• tests/js/demo_merge_parity.mjs (renamed from demo_fusion_parity.mjs, checks fusion.js or songs.js by the manifest entry): 16/16 settings identical for Just-Bird and for All-In-One
• Headless Chromium, built site: Just-Bird → All-In-One → Just-Bird on the sample (55, 29, 55 boxes), species column, song gap slider, JSON export, phone width, dark mode; no console errors
• check_license_headers.py passes
• Real browsers on macOS / iOS with All-In-One (about 3 times slower than Just-Bird: 31 s for the 2-minute sample in headless Chromium)

docs/models/All-In-One_fp16.onnx adds 51 MB to the repository.

🤖 Generated with Claude Code