laion/vocal-burst-synthesis-listening

Six burst classes, each re-trained on synthesised same-speaker data. frustrated_groan produced a detector hit rate of 0.000 at every merge weight — its adapter had been trained on 568 rows of which only 4 contained the burst at all. This page lets you hear an adapter re-trained on ~2,000 synthesised rows against the original.

The measurement cannot settle it: the detector names the source dataset's own label 3.4 % of the time, and the label it gives the new adapter's output — Exhausted Groan — is the same label it gives the real, clean Frustrated Groans the training data was built from. So the strict score stays at 0.000 while the family-relaxed score moves 0.000 → 0.259. The ear is the arbiter here.