Pick a set with the tabs; hover a system to isolate it, click to hide it; the threshold slider moves the ● and the F1 column; the length filter restricts curves and markers to utterances of the chosen lengths.
our models shown:
off: operating points onlyThreshold transferability: one setting, every set
Production requirementAttribution precision ≥ 95 %so that other people's speech is not attributed to the wearer
Compare with (pick 2)
One threshold per system, every set. Each system starts at the lowest threshold that holds precision at or above 95 % on every set, and the cards show what that single setting costs in recall, set by set. A system that passes here can ship with one threshold; one that fails needs calibration per deployment.