Canopy sound model for the in-browser demo on /how-it-works/
=============================================================

panns-cnn14-w8-v1.onnx.part0-3   (join in order: 87,523,243 bytes,
                                  sha256 0a74101922414a8036cc0f077fbbc8384bea4b482bf0f8b796c8508881828e76)
  PANNs CNN14 audio embedder, as served by the Canopy API, with conv blocks
  4-6 and fc1 stored as 8-bit weights (per output channel) and dequantised at
  run time; the front end and conv blocks 1-3 stay float32. Embedding cosine
  against the float model >= 0.9998 on field audio.
  Source weights: Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang
  and Mark D. Plumbley, "PANNs: Large-Scale Pretrained Audio Neural Networks
  for Audio Pattern Recognition", pretrained models, Zenodo,
  https://doi.org/10.5281/zenodo.3987831, licensed CC BY 4.0
  (https://creativecommons.org/licenses/by/4.0/). Changes by Canopy: exported
  to ONNX, AudioSet output layer split out, conv blocks 4-6 and fc1
  compressed to 8 bit.

audioset-head-v1.bin
  The AudioSet output layer (fc_audioset) of the same checkpoint:
  float32 weight [527 x 2048] row-major, then float32 bias [527]. CC BY 4.0, as above.

audioset-labels-v1.json
  AudioSet class names, from the AudioSet ontology by Google Inc.
  (https://github.com/audioset/ontology), licensed CC BY-SA 4.0
  (https://creativecommons.org/licenses/by-sa/4.0/). Unchanged.

samples/*.wav
  20 s excerpts of Canopy's own field recording at Beri, Rajasthan,
  18 Oct 2026, from the minutes in their names (IST), mono 32 kHz.
  Screened automatically for speech.

manifest.json
  Files, hashes, rows and alert lines read by /how-it-works/try.js. Written by
  scripts/export_site_try_model.py.

The browser runtime is ONNX Runtime Web 1.30.0 (MIT); its licence and the
notices for code compiled into it are in /vendor/ort/NOTICE.txt and
/vendor/ort/ThirdPartyNotices.txt.