ASPECT-D · v1.4

A headline that was a property of the coordinate

The v1.0 primary result was an exponent contrast, Δτ = τ_WER − τ_SIM. Refitting the identical measurements after a monotone change of the error variable sends it negative. This page records what failed, what replaced it, and what independently survived.

What changed Δτ is invariant to affine rescalings of the error — which the pre-registration checked — but not to monotone non-affine ones. It also halves each time the fitted T window doubles. The paper now leads with a model-free, floor-referenced contrast that moves by neither.

1The failure: coordinate

Same 225 surface points, same fitting code, same weights, same 32 multi-starts. Only the error variable changes.

error coordinate	τ_WER	τ_SIM	Δτ	verdict
identity (pre-registered)	0.8362	0.7260	+0.1102	as published
log(1+err)	0.6665	0.6913	-0.0248	sign flipped
sqrt(err)	0.4325	0.6842	-0.2517	sign flipped
err capped at 1	0.7093	0.7260	-0.0167	sign flipped
err squared	1.5613	0.8122	+0.7491	as published
affine 3+2*err (v1.0 check)	0.7600	0.7249	+0.0351	as published

2The failure: range

We extended all 45 runs to T=64 at the full 400 items — 315 balanced rows, no config or item subset. The earlier extended fit used 7 of 15 configs at 200 of 400 items, so it confounded range with subset. Refit on nested windows:

window	Δτ	rows
T<=8	+0.2191	180
T<=16	+0.1102	225
T<=32	+0.0603	270
T<=64	+0.0660	315

retracted as a headline Δτ roughly halves each time the window doubles, and three of five reparameterisations flip its sign. It is reported in the paper as a pre-registered test that passes in its declared coordinate and range, with both conditions attached.

3The replacement

Referenced to floors we measure — ASR word error on the real recordings (0.0345) and the speaker similarity of a codec round trip (0.5554) — rather than to zero. No model is fitted.

T	intelligibility closed	identity closed	ratio
2	32.2% [29.7, 34.4]	20.0% [19.1, 20.9]	1.61×
4	69.8% [66.0, 73.1]	35.7% [34.3, 37.0]	1.96×
8	81.8% [78.3, 84.8]	43.5% [42.1, 44.9]	1.88×
16	86.2% [82.8, 89.1]	46.4% [45.0, 47.8]	1.86×
32	87.9% [84.9, 90.5]	47.5% [46.2, 48.8]	1.85×
64	88.4% [85.3, 91.0]	47.9% [46.6, 49.2]	1.85×

headline At the deployable T=16 default, refinement closes 86.2% of the reachable intelligibility range and 46.4% of the reachable identity range. The ratio holds between 1.85× and 1.96× across a 16× range of refinement budget, and between 1.68× and 1.86× across the coordinates that flip Δτ's sign — because numerator and denominator move together.

Identity plateaus: quadrupling the budget to T=64 takes intelligibility to 88.4% but identity only to 47.9%. More refinement is not the lever, which is the paper's case for spending inference compute on search instead.

4What independently survived

The paper's top stated limitation was that selector and scorer shared the ECAPA-TDNN lineage. Five encoders across four families rescored the same clips. Gate G0(c) rejects three of them on absolute cosine scale despite AUC ≥ 0.983, so agreement is read from per-item win rates, which are invariant to any monotone rescaling of an encoder's cosine.

encoder	role	win rate	95% CI	raw Δ
wavlm-large	selector (circular)	84.9%	[83.6, 86.1]	+0.0803
xvect	independent family	79.0%	[77.5, 80.5]	+0.0075
ecapa	gated scorer	72.3%	[70.6, 73.8]	+0.0444
ge2e	independent family	69.1%	[67.5, 70.8]	+0.0164
wavlm-base-plus	independent scale	64.6%	[62.9, 66.3]	+0.0173

closed 4/4 non-selector encoders put search above refinement with every interval clear of one half, across 15 runs. The raw Δ column is scale-bound and not comparable across rows; the win rate is.

5Three further scope conditions now in the paper

condition	measurement	value
the declared form is rejected by its own criterion	χ²/dof, WER / SIM-o	14.3 / 8.8
the declared interval omits weighting uncertainty	Δτ CI with weights resampled	[0.0349, 0.1640]
no global depth–step exchange rate	local κ = τ/β implied by the separable fit	0.41

The fit is kept as a compact description of curvature. No mechanism is read from its parameters.

6Reproduce

bash src/finish_v14.sh collects the extended surface, refits, regenerates every number on this page and in the paper, rebuilds, and verifies. It guards the v1.0 immutables against the release tag, and src/paper_v14.py refuses to emit the range macros when the fitted windows are identical rather than shipping a tautology. python src/check_claims.py fails the build if a retracted claim is live anywhere.

← project page · v1.1 extensions · results · protocol