I Recorded the Path a Real Voice Takes. Six Representations Could Not See It, So My Two Hypotheses Were Never Tested.

A trained singer sliding continuously from clean phonation into rasp — 58 recorded traversals, every intermediate point a real larynx rather than a model's guess. A validity gate written before any embedding existed asked whether the instruments could resolve the path at all. Across three speaker encoders and three neural codecs, at four window settings and every scope computed, it returned SNR 0.53 to 1.85 against a required 3.0. Zero cells passed. The two hypotheses that motivated the whole study were therefore never computed, and they are recorded as unreachable with these instruments — not as refuted.

The assumption nobody has checked against a recorded path

Voice conversion, attribute editing and speaker-blend conditioning all do the same geometric thing. They draw a straight line between a source point and a target point in a learned representation, and treat the coordinates in between as natural intermediate voices. kNN-VC blends WavLM features. Speaker blending interpolates embeddings. Attribute editing displaces a latent along a direction vector.

The premise is not unreasonable. It is untested, and it is untested for a structural reason: in every one of those settings the path is synthesised. The endpoints are real and everything between them is a product of the model being evaluated. There is no recording of what a voice actually does on the way from clean to raspy, so there is nothing to compare the interpolation against.

Two independent bodies of work predict the real path is neither straight nor symmetric. Rasp and growl recruit secondary laryngeal structures — ventricular folds, aryepiglottic constriction — at thresholds rather than proportionally, and a threshold-recruited articulator does not trace a linear acoustic trajectory. Separately, phonation threshold pressure is well established to differ between oscillation onset and offset, which is hysteresis in the vocal folds themselves and settled physiology rather than an open question. Neither is a claim this study makes. The open question was whether learned speech representations reflect either property.

So I recorded the path

A professionally trained speaker sustains a vowel and slides continuously from clean phonation into rasp, growl or breathiness and back, and from low pitch to high and back. 58 traversals, 2.2 to 8.5 seconds each. Ordinary corpora contain the endpoints of these transformations. This one contains the trajectory between them, which makes two otherwise unanswerable questions answerable: whether the real human path is straight in these spaces, and whether it is the same in both directions.

Both questions were pre-registered on 14 August 2026, before any embedding or latent was computed — together with a validity gate that had the power to stop the study.

A check that cannot fail is not a check

The gate exists because of an earlier failure of my own worth naming. Study 2's validation gate flagged 0 of 233 utterances, because its bounds were set from population plausibility: wide enough to be uncontroversial, and therefore too wide to discriminate anything. Study 3's gate is set from this project's own steady recordings instead.

The logic is simple. Study 1 had already measured that sustained vowels sit near the encoder floor. If a windowed sustained-vowel embedding is noise-dominated, then a curvature statistic computed on a traversal is measuring window-to-window jitter and reporting it as path shape. So, for each representation and window setting: window the steady clean vowels and measure how far the representation moves while the voice is not changing; measure the endpoint-to-endpoint distance of each traversal; divide.

SNR = median chord length ÷ 95th percentile of within-steady scatter. Pre-specified criterion: SNR ≥ 3.0 in at least two of three Tier-A representations. The failure branch was fixed in advance — fail in Tier A, abandon speaker embeddings and proceed to codec latents; fail there too, and publish the gate itself as a measurement result. The gate result was pre-registered to be published whatever it showed.

Tier A — zero of three, at every window length, in both scopes

Every cell sits near 1.0 against a requirement of 3.0, and an SNR of roughly 1.0 has a direct reading: the endpoint-to-endpoint distance of a real vocal traversal is no larger than the first-to-last distance measured within a steady sustained vowel in which nothing is changing.

WavLM cannot embed a 250 ms window at all — its x-vector TDNN head needs a longer input than the window provides. The whole setting is declared unavailable rather than partially filled, because dropping the failing files and computing a path from the remainder would bias the result toward whichever files happened to survive. That is a capability limit of the model, not an error in the pipeline.

The instrument is verified working

A failed gate is only informative if the apparatus is known to function, so two controls ran before the failure was accepted.

The first one inverted, and it is published as a mis-designed control rather than deleted. It assumed connected speech would be more self-consistent than sustained vowels through the identical code path. The opposite held — median within-file cosine 0.205 for steady connected speech against 0.440 for sustained vowels. The reason is obvious in hindsight: at 500 ms, different windows of running speech contain different phonemes, so low self-consistency is expected and says nothing about pipeline integrity.

The second is decisive. Within-speaker window pairs against between-speaker pairs on matched material:

Separation is positive throughout and increases with window length in every encoder. The representations do carry speaker identity on this material. The instrument works; the resolution does not suffice.

The obvious artifact, checked and excluded

The first and last window of any recording contains phonation attack and decay, which are acoustically extreme. If those dominated both the steady and the traversal measurements they would produce exactly the observed SNR near 1 — as an artifact of the gate's own design rather than a property of the representations. Recomputed with endpoints taken at 20% and 80% of each file:

Largest movement is 0.39 and no cell approaches 3.0. The negative is verified rather than merely observed.

Tier C — codec latents fail too, and at the native frame rate they fail worst

This was the pre-registered branch, and the reason it was named in advance is that codecs are frame-level, so the windowing constraint that governs Tier A does not apply to them. Three families loaded and ran: EnCodec 24 kHz, DAC 44 kHz, Mimi. Distances are raw Euclidean — SNR is a ratio so scale cancels, and no normalisation assumption is imposed on latents that are not natively normalised. L2-normalised figures agree to within 0.02 throughout.

The frame-level advantage over Tier A is real and it is irrelevant. More points on the path do not help when each point's noise is the size of the whole signal.

Overall: six representations, four window settings, zero passes, SNR range 0.53 to 1.85 against a required 3.0. That range spans every speaker scope computed; restricted to the two scopes tabulated above it is 0.53 to 1.67, and the maximum of 1.85 is a secondary-speaker cell reported separately and never pooled. No cell, in any representation, at any window length, in any scope, reaches 2.0.

What the failure actually says

The per-window scatter of a steady sustained vowel is the same order of magnitude as the displacement produced by a complete vocal transformation. Path shape cannot be recovered from these representations on this material, because any curvature estimate would be dominated by window-to-window jitter and would report that jitter as geometry.

The trend across window lengths names the tension precisely, and it is not one that more careful engineering resolves. Discrimination improves monotonically from 250 to 500 to 1000 ms, so longer windows would indeed fix the noise — but the traversals are 2.2 to 8.5 seconds long. At the two to three seconds these encoders want in order to produce a stable embedding, a traversal yields only one to four windows, and one to four points is not a path.

Speaker encoders require more audio to produce a stable embedding than a continuous vocal traversal can spare while still resolving its own shape.

That is a measurement statement about the instruments. It is quantified, it holds across three encoder architectures and three codec families, and it holds at every window length tested.

The status of the two hypotheses, which is the part most likely to be misreported

H1 — that the real path deviates from the straight chord — and H2 — that forward and reverse paths differ at matched acoustic progress — were never computed. No statistic bearing on either one exists.

They are unreachable with these instruments, not refuted, and the distinction is not a formality. A refutation would be a claim about vocal traversals: that they are straight, that they are symmetric. What was actually established is a claim about ECAPA-TDNN, ResNet, WavLM, EnCodec, DAC and Mimi — that their resolution on this material is insufficient to answer the question either way. Any citation or summary that converts "unreachable" into "falsified" is misreporting this result.

One descriptive observation, explicitly outside the hypothesis frame

At 500 ms under ECAPA, the median traversal chord is 1.20 in Euclidean distance — cosine about 0.28 between a traversal's start and its end. Two different speakers, a man and a woman, producing the same vowel in the same condition, sit at cosine 0.555.

A single speaker sliding from clean phonation into rasp moves further through this embedding space than the distance between two different people.

Three caveats are attached at birth rather than in response to review: the secondary speaker's recording channel is unverified, sustained vowels are known to be weak material for speaker identity, and the same window noise that failed the gate applies to both numbers. It is not evidence for H1 or H2, which concern the shape of the path rather than its length.

Read with those caveats intact, it still bears on speaker-verification practice. A verification threshold is calibrated on the assumption that within-speaker variation is small relative to between-speaker variation. On this material, under this encoder, at this window length, a deliberate change of vocal quality by one person violates that ordering. This study cannot establish how often that happens in deployment — but it does establish that the ordering is not guaranteed by the geometry.

What this does not establish

• H1 and H2 are not rejected. They were not computed. They remain open.
• Nothing here says continuous vocal attributes do not exist, or that their paths are straight, or that they are symmetric.
• Speaker encoders and neural codecs are not shown to be invalid in general. The claim is bounded to path resolution, on this material, at traversal timescales. Control 2 shows these same encoders separating speakers on the same audio.
• Two speakers, one fully controlled, roughly five minutes of audio. This is a case series about representations, not a population claim about voices. No result here generalises across speakers.
• Sustained vowels only. Whether any of this transfers to connected speech is untested and not claimed — and Control 1 is a direct warning that connected speech behaves differently under identical windowing.
• The 3.0 criterion is an analysis choice, declared as such in advance. A lower threshold would not rescue the study: the observed maximum anywhere is 1.85.
• No priority or novelty claim is made anywhere. A literature scan was obtained and is treated as a list of leads, explicitly not as an audit — it returned no identifiers for most entries and at least one of its citations could not be confirmed to exist. Absence of a search result is not evidence of absence.

What a future study would need

Stated so the failure is usable rather than merely discouraging. The gate failed on a resolution mismatch with a specific shape — stable embeddings need seconds, and a traversal has seconds in total — and none of the three ways out is a refinement of what was run here.

Longer traversals. Execute the transformation over 15 to 30 seconds rather than 2 to 3, and a 2 s window yields enough points to constitute a path. That is a recording decision, not an analysis decision, and it costs a session rather than a compute budget. It is the cheapest of the three.

Representations built for short spans. Everything tested here is trained for utterance-level or frame-level reconstruction, not for stable short-span identity-and-quality coordinates. A representation whose stability at 250 ms is a design target would be evaluated on a different footing.

A different observable. Direct acoustic parameterisation — HNR, CPPS, F0 — is measurable on this same audio at frame rate and does not inherit the encoder's noise floor. It answers a narrower question, and it answers it.

The dataset remains what it was before the gate ran: a few hundred labelled continuous traversals, which is an evaluation benchmark rather than a training corpus — a way to score whether a model's latent path matches a real one, which the disentanglement literature does not otherwise have. What this study establishes is that scoring it through current speaker encoders and codecs, at traversal timescales, is not yet possible.

Why publish a study that stopped itself

The gate was written before the data existed. Its threshold was set from this project's own recordings rather than from population bounds wide enough to be uncontroversial. Its failure branch was named in advance, and its terminal outcome — this paper — was named in advance as well. The pre-registration also froze the analysis set, and specifically forbade the moves that would have rescued a null: adding a progress variable after seeing it, re-running at an unlisted window length, introducing dynamic time warping after a failed test, pooling the two speakers. None of those was done.

So this is not a disappointment retrofitted into a finding. It is the pre-declared outcome of a check built to be capable of failing, run on an instrument verified to work, with the most likely artifact explicitly excluded.

Paper: Continuous vocal-attribute traversals fall below the path resolution of current speech representations: a pre-registered validity gate across three speaker encoders and three neural codecs — Zenodo, 6 September 2026, CC BY 4.0.

DOI: 10.5281/zenodo.22538235 — the concept DOI, which always resolves to the current version. Study 3 of the BedVibe vocal-representation series; Studies 1 and 2 are at 10.5281/zenodo.21921958.

Every figure in the paper is derivable from the tables it releases, without access to the audio. The recordings themselves are not released with this study — they are rights-cleared, consented and covered by signed talent contracts, and their release is a separate decision.