A style vector is a mean in embedding space, and rough audio does not sit near the middle of that space [8]. It sits far out along the hoarseness direction, so four clean clips plus one outlier produce a centroid pulled well toward the outlier [8]. That geometry is why the defect showed up in every sentence instead of one in five, and it is why the roughness of a single file propagates to every output of that style [9]. Weigh the two paths against each other: in a corpus of about 200 clips one bad take is 0.5% of the training signal [6], while in a five-clip registration the same file is 20% of the exemplar set [7], roughly 40 times the influence for an identical artifact [1].
The admission test is the part worth staring at. Script matching is a text comparison standing in for an acoustic judgement, so a clip can crack an octave mid-sentence and still be scored as correct because the words are all there [5][10].
The replacement metric is small and specific. It extracts an F0 series every 10 ms, keeps voiced frames, and computes jitter as the mean absolute difference between adjacent periods divided by the mean period [11]. Octave jumps are counted as the fraction of frame transitions where the F0 ratio exceeds 1.8 or falls below 0.55 [12]. Candidates are sorted by jitter plus twice the jump rate, lowest first [13], with jumps weighted more because a discrete break is more audible than continuous roughness [14]. Note that 1/1.8 is 0.5556, so a drop to 0.554 of the previous frame escapes the down-jump test while its upward mirror at 1.805 gets caught [4]. Clips with fewer than 10 voiced frames return None and drop out of the candidate pool [16], which at one frame per 10 ms means anything under 100 ms of total voiced audio is simply unjudgeable [2]. That guard exists because short interjections would otherwise score as the most stable clips by chance [16].
Now treat the reported numbers as claims about someone else's recordings. Autocorrelation F0 estimation mistakes harmonics for the fundamental on lower voices, and those errors land in the jitter term [18]. On the male voices here, jitter read 31% to 56% against a physiological norm below 1%, so measurement error dominates the value [19]. Two consequences follow from the arithmetic. The jump term only outweighs the jitter term when the jump rate exceeds half the jitter, which in that range means 15.5% to 28% of frame transitions [3]; the nominal 2.0 weight does less sorting work than it looks like it does. And the ranking survives only because the error is close to uniform across clips from one speaker [20]. A constant offset does not change sort order, but the variance of that error across clips does, and the post reports the magnitude without reporting the spread.
For the number to transfer you need exemplar averaging in embedding space, a transcript-only gate ahead of it, one speaker per style [21], and an autocorrelation pitch tracker. Swap in a neural F0 estimator and the thresholds move; the shape stays, because when k is five, per-item quality is a first-order term.
