DGE all-tags tagger (combined)
The sound tagger of Document Graph Explorer: EfficientAT
networks fine-tuned by scripts/audio-model/, combined by combine.py so each of the app's tags comes from the run that
scored best on validation clips (the held-out sets are scored once, after choosing). Runs: all-tags, all-tags-run5, all-tags-run7, all-tags-run8, all-tags-run9, all-tags-run11, all-tags-run10, all-tags-recall2, all-tags-r29, all-tags-r31, all-tags-r32, all-tags-r33. picks.json lists every tag's choice and validation numbers,
eval.txt the combined model's held-out scores. Input: 10 s of 32 kHz mono (samples32k); output: scores (sigmoid).
Base networks and AudioSet weights: EfficientAT (Florian Schmid et al.), MIT.
Licence: CC BY-NC-SA 4.0, non-commercial only. One of the runs was trained on Mixing Secrets multitrack songs (RawStems), whose licence allows non-commercial research and education only.