Neural Reconstruction Codecs Can Mask Pathological Voice Jitter: A Reference-Relative Analysis of Twelve Codecs

Remote voice assessment reads clinical acoustic indices off a decoded waveform, so the codec sits inside the measurement chain. We audit twelve codecs at their deployed operating points, five reconstruction neural and seven deployed-classical, as perturbation operators before a local-jitter estimator, on 485 pathological and 687 healthy vowels (Saarbrücken Voice Database, SVD) and 154 dysphonic and 107 normal-sounding vowels (PVQD). The pre-specified estimand is reference-relative, the change Δ|z| in standardised distance from the healthy centre, because a naive z-difference cancels the healthy location. Holding this curated panel fixed and resampling speaker clusters, all five reconstruction estimates are negative on both corpora and the pole mean excludes zero (−0.111, 95% CI [−0.166, −0.055] on SVD; −0.213 [−0.270, −0.151] on PVQD, where the poles separate exactly at permutation p = 1/792). The raw pathological-minus-healthy interaction of −0.333 is largely baseline geometry: conditioning on where a reading already sits relative to the healthy centre removes most of it, leaving a residual of −0.13 to −0.16 where the groups overlap on SVD. The seven deployed-classical arms stay inside the estimator's repeatability floor. The effect is bound to the operating point, and the audit is fragile to provenance: one encoder binary swap manufactured a spurious +0.50 classical effect that a unified re-run reduced to +0.015.