moves the ● on Figures 3, 5 and 6 together
65%72%79%86%93%100% 05%10%15%20%25% frames emitted under a wrong enrollment (absent-speaker false accepts) recall of the wearer's speech Isolate 1.0Isolate 1.1Isolate 1.1 + SBM-1Isolate 2.0Isolate 2.0 + SBM-1 ● = selected threshold MEASURED · enacted-retail set · 90 s blocks
Recall of the wearer's speech against the absent-speaker false-accept rate as the threshold sweeps, enacted-retail set. The whole curve moved up and to the left between generations, so at any leakage budget the current model keeps more of the wearer's speech. This is the curve a deployment chooses from: fix the leakage budget, read off the threshold.