I'm training text2asmr, a text-to-ASMR model (whispering plus trigger sounds like tapping, crinkling, brushing). The generator needs every training clip tagged with the sounds in it, and there are about 126k recordings, so I tried frontier models as labellers. They were much worse than I expected, so here are the numbers.

Setup. Every judge got the same prompt and the same 20 labels on 480 clips:

External: sound-effect clips from FSD50K and ESC-50 that humans had already labelled with one class, 40 per class. This measures recall.

People: 6-second ASMR clips labelled on our site. A class counts as present when at least half of a clip's labellers ticked it.…