[El1as42]

Summary

• Streamlit NMS IoU 0.5 → 0.7. The CLI default and the ONNX graph both use 0.7, and the window fusion constants were tuned with it. On 13 test recordings, 0.5 alone changed about 3% of the Just-Bird boxes.
• ONNX demo accepts up to 10 minutes of audio (was 60 s), the same limit as the Streamlit demo. The ONNX graph has no length limit of its own: it walks the file in 60 s PCEN segments inside an ONNX Loop.
• Audio is resampled once. It is decoded straight into a 32 kHz OfflineAudioContext. Before, the default AudioContext resampled to the hardware rate (often 48 kHz) first.
• The demo sends unpadded audio to the graph. The graph pads to the clip grid itself and trims detections before window fusion, as detect_birds.py does.
• The spectrogram is drawn as 60 s canvas tiles. A single canvas broke at about 5.5 min (browser limit of about 32k px).
• Removed a duplicate write_metadata in onnx_export.py.

Parity check

detect_birds.py with models/Just-Bird.pt (conf 0.18, phrase gap 0.05) against the all-in-one graph, run in Python ONNX Runtime. Audio: tests/test.wav plus the 12 parity soundscapes, 1231 reference boxes. A box counts as matched at IoU ≥ 0.5.

Compared with detect_birds.py (NMS 0.7)	Matched	Mean IoU
ONNX fp32	99.7%	0.999
ONNX fp16 (docs/models)	99.6%	0.995
Streamlit before this PR (NMS 0.5)	97.3%	0.965

Browser demo in headless Firefox, compared with Python ONNX on librosa-resampled audio (60 s excerpts):

Input	Old page (double resample)	This PR
32 kHz	16 boxes vs 15, mean IoU 0.993	15 vs 15, mean IoU 0.998
48 kHz	40/42 matched, 0.933	41/42, 0.970
44.1 kHz	16/18 matched, 0.935	16/18, 0.969

For input that isn't 32 kHz, the remaining gap comes from the browser's resampler. The page now says so.

Test plan

• mkdocs build, then the demo page in headless Firefox on 300 s (32 kHz) and 200 s (48 kHz) files: 104 and 129 boxes, 5 and 4 spectrogram tiles, 25 s and 15 s of inference
• The 300 s run matches Python ONNX: 104/104 boxes, mean IoU 0.996
• Screenshot across a tile edge: no visible seam, and boxes and labels continue across it
• onnx_export.py compiles
• Full 10-minute files (11 min input, cut to 600 s), 32 kHz mono and 48 kHz stereo:
  • Firefox 136 (headless): 344 and 354 boxes in 48 s and 53 s; 10 spectrogram tiles; largest process peaked at 2.2 GB, about 1.1 GB during inference
  • Chromium 156 (Playwright): 344 and 357 boxes in about 47 s each; browser process about 1.1 GB during inference
  • All three engines give 344/344 boxes against Python ONNX on the 32 kHz file (mean IoU 0.998)
• WebKit 18.2 (Playwright's frozen Ubuntu 20.04 build, the closest to Safari on Linux): results are correct (344/344 on 600 s), but its web process used 4.9 GB on a 60 s file and up to 12 GB on 600 s, and one 600 s run crashed the page. The page before this PR uses the same 4.9 GB on the same 60 s file, so this predates the PR and depends on ONNX Runtime Web in this WebKit build, not on the audio length
• Real Safari on macOS / iOS (not available on Linux); WebKit's memory use makes this the one to check

🤖 Generated with Claude Code