Tritium autoradiography maps where hydrogen hides inside steel. Each radioactive decay exposes one silver grain in a photographic film, and that grain marks the decay only to within about 150 nm. A dislocation core is under a nanometer wide, so no single grain can say whether the hydrogen sat in the core or in the strained region around it. Singing voice synthesis has a close cousin of this problem that it has not framed this way. Modern systems compress singing into discrete tokens, short learned codes, here emitted 50 times per second. A second generator then "refines" those tokens into a detailed spectrogram. That refinement stage visibly adds vibrato depth, pitch glides and sharp note onsets. The open question is whether the added detail is the singer's actual detail, recovered from what the tokens encode, or a plausible detail sampled from the generator's training prior.
We propose to measure this directly. The idea is to treat the token stream the way autoradiographers treat the silver grain: as a detector with a finite resolution. We then ask, at each timescale of pitch movement, whether the refined output is determined by its input or merely consistent with it. The deliverable is a resolution-limit curve for token-based singing synthesis, the analogue of the 150 nm figure. It is paired with a simple fix that spends continuous side information only on the timescales that fall below that limit. As a secondary check, the same protocol is run on simulated autoradiographs to test what sub-resolution claims in the hydrogen literature can actually support.
Core Claim
Detail finer than a detector's resolution can be generated but not recovered. A good generator can produce such detail, and ordinary quality metrics will reward it anyway. The test that separates the two cases is simple. When refinement truly recovers detail, two independent samples from the refiner agree with each other much more closely than either agrees with a random plausible alternative. When it invents detail, the real recording behaves like just another sample. We therefore define, for a band $b$ of pitch-contour timescales (note-level, vibrato at 5–7 Hz, transitions of tens of milliseconds), a determination ratio
$$D_b = \frac{\mathbb{E}\, d_b(\hat f^{(1)}, \hat f^{(2)})}{\mathbb{E}\, d_b(\hat f, f)},$$
where $\hat f^{(1)}, \hat f^{(2)}$ are two refiner samples from the same tokens, $f$ is the ground-truth pitch contour, and $d_b$ is the mean squared difference after band-pass filtering to band $b$. If the truth is exchangeable with the samples, $D_b \approx 1$. If the tokens pin the detail down, samples collapse around the truth and $D_b \to 0$. The resolution limit is the finest band at which $D_b$ stays below a fixed threshold, such as 0.5. We predict three things. First, that the limit sits near the transition scale rather than the vibrato scale, because 20 ms frames can carry a 150–200 ms vibrato period but not the fine shape of an onset glide. Second, that codebook quantization erodes vibrato depth even where the frame rate suffices. Third, that aggregate pitch error hides both effects.
Grounding in Prior Work
The motivation comes from the singing side. Adapting Speech Language Model to Singing Voice Synthesis adapts a 1.7B speech language model to singing using 50 fps score and codec tokens. It adds a conditional flow matching stage, a generator that learns a smooth path from noise to data, which turns tokens into mel-spectrograms with explicit pitch re-injected. Their ablation shows the refinement stage matters for perceived quality. Their pitch error (62.79 F0-RMSE) is nonetheless worse than the token baseline TokSing (55.83) at equal SingMOS of 4.09. That pattern is exactly what one would expect if refinement adds convincing detail that is not the reference's detail. CoMelSinger documents a related failure, prosody leakage, in which a prompt's pitch bleeds into the output. It is further evidence that discrete tokens carry entangled, partially lost pitch information. StylePitcher generates expressive pitch curves by rectified flow matching and reports that an LSTM classifier distinguishes them from real curves at roughly 50%. That result is a realism test, not a fidelity test, and our metric is designed to supply the missing fidelity test.
The resolution framing comes from the autoradiography side. Analysis of Hydrogen Behavior in Cr-Mo Steels by Electron Microscopic Autoradiography states a resolution of about 0.15 μm. The same paper reports that over 50% of silver grains lie within 30 nm of martensite lath boundaries. A statistic at 30 nm drawn from a 150 nm detector can only be a population-level inference, never a per-grain localization. The distinction between what a detector resolves and what an ensemble statistic can support is the lens we carry into singing. Observation of Hydrogen Trapping Sites at 77 K Exposure in SUS316L Steel acts as a constraint: freezing the specimen during exposure removes diffusion blur but leaves the grain-size limit intact. In the same way, a cleaner codec or better training cannot remove a frame-rate limit.
The theory comes from Computational resolution limit: a theory towards super-resolution and Learning Point Spread Function Invertibility Assessment for Image Deconvolution. Both show that recoverability collapses beyond a noise-dependent cutoff and that the cutoff can be estimated. We borrow the concept but replace an analytic blur model with an empirical, sample-based test. A learned codec has no closed-form blur, so an empirical test is the only one available.
Research Program
The first stage is measurement. We run a resynthesis protocol built on Muskits-ESPnet: real singing is encoded to tokens and decoded through several refiners, with multiple samples drawn per utterance. Pitch is extracted from every output and decomposed into timescale bands with a wavelet filter bank. We then compute $D_b$ and the band-wise error to truth. We vary three quantities: token frame rate (by retraining the tokenizer at 25, 50 and 100 fps), codebook size, and refiner type. Varying them lets us tell which factor sets the limit in each band.
The second stage is a targeted fix. Suppose the limit lands at transition timescales, as predicted. We then add a sparse continuous side channel to the refiner's conditioning. It carries the band-limited pitch residual only in bands where $D_b$ exceeds threshold, at a bit budget far below full-rate F0. We compare this against full continuous pitch re-injection at the same total conditioning budget.
The third stage is a transfer check. It stays small and relies only on simulation, since no digitized autoradiograph dataset is available. We simulate tritium sources placed either at lath-boundary cores or in a 1/r strain field around them. Each source is spread by a beta-range kernel and assigned by Poisson draws to 100–200 nm grains, and $D$ is computed for a learned deconvolver. The question is how many grains are needed before core versus strain-field trapping becomes determined rather than invented. This tests whether the protocol is a general tool for generative inversion under a coarse detector. It is not the core contribution.
Baselines
The naive reference point is the codec decoder with no refinement, applied directly to the tokens. It shows what the tokens alone determine and should give low $D_b$ wherever it outputs detail at all, because it adds no sampled detail. Beating it means achieving lower band-wise error to truth, not just higher mean opinion score.
The strongest existing method is the token-plus-conditional-flow-matching pipeline of Adapting Speech Language Model to Singing Voice Synthesis, which reports SingMOS 4.09 and F0-RMSE 62.79 on the ACE-Opencpop/ACE-KiSing benchmark. TokSing on the same benchmark reports SingMOS 4.09 and F0-RMSE 55.83, and XiaoiceSing reports 3.88 and 71.67. We retrain this refiner in Muskits-ESPnet on the same data splits. Every variant gets the same training steps and the same hyperparameter search budget, and each is sampled five times per utterance. On the Opencpop corpus alone, the CpopSing baseline reports MOS 3.70 and F0-RMSE 27.58 (Opencpop). We cite it as a scale reference only, since the setups differ. Our fix counts as beating the strongest method if it lowers transition-band and vibrato-band error to truth without lowering SingMOS below 4.09 under our re-run, and if it drives $D_b$ below threshold in those bands.
The key ablation isolates band selection. We compare full-rate continuous pitch re-injection against our sparse residual channel at matched bits. If full re-injection matches ours at equal budget, the band-selective part adds nothing and the contribution reduces to the measurement protocol. The file contains no reported number for the refinement-free ablation in the Zhao et al. paper beyond the qualitative drop, so we measure it ourselves.
Why not just…?
Why not just report F0-RMSE and MOS, as everyone does? Aggregate pitch error mixes all timescales and treats invented detail and a smooth guess alike. It can even penalize a correct vibrato that is slightly out of phase. The benchmark numbers above show the problem: two systems tie on SingMOS but differ by seven F0-RMSE points, and nothing in those metrics says which timescales account for the gap. The determination ratio is the only quantity here that separates "matches the reference" from "matches the prior." Experiment one checks directly whether it ranks systems differently from F0-RMSE.
Why not just raise the token frame rate or codebook size? This is a real threat, and it may fix the problem outright. The frame-rate sweep is designed to settle it. If 100 fps tokens bring every band below threshold at acceptable cost, we will report that as the answer, with the resolution curve serving as the evidence for the needed rate. The side channel then matters only if the rate needed for transitions is much higher than the rate needed for everything else, which is what the autoradiography analogy predicts, since grain size is fixed while trap geometry varies.
Why not just feed continuous pitch everywhere? Zhao et al. already re-inject pitch. Their pipeline still ends up with higher pitch error than TokSing, so re-injection alone does not guarantee fidelity. The matched-budget ablation tests whether spending bits uniformly is wasteful compared with spending them where $D_b$ says tokens fail.
Why not skip generative inversion in the autoradiography check and fit a parametric likelihood? For two clean hypotheses, a Poisson likelihood with a known beta-range kernel is probably optimal, and we use it as the reference. The learned deconvolver is only worth keeping if it matches that likelihood's decision on clean cases. That makes the simulation a calibration test of the protocol, not a claim that flow matching beats physics.
Proposed Experiments
The first decisive check runs on GTSinger (dataset, paper). GTSinger pairs natural and technique-emphasized takes of the same material and annotates vibrato and glissando at the phoneme level. Computing $D_b$ separately on annotated vibrato and glissando segments lets us test the prediction that glissando shape is invented while vibrato rate is recovered. The paired takes add a sharper test. If the refiner's vibrato depth on the emphasized take is no closer to that take's truth than to the natural take's truth, depth information is not surviving the tokens.
The second check uses Opencpop's slur labels, where one vowel spans several notes. Slurs are exactly where transition detail should matter, so they test whether the limit concentrates at note boundaries.
The third check reproduces the ACE-Opencpop/ACE-KiSing comparison under our re-run and adds band-wise errors and $D_b$ to the reported SingMOS and F0-RMSE. It asks whether the TokSing-versus-refiner gap in F0-RMSE lives in the transition band.
The fourth check is the matched-budget side-channel ablation across the 25/50/100 fps tokenizers. The fifth is the simulated autoradiography calibration. It reports how many grains per boundary are needed for core versus strain-field placement to become determined, and whether a 30 nm proximity statistic like the SCM435 one can be distinguished from a uniform 150 nm-scale spread at realistic grain counts.
The idea fails if $D_b$ stays near zero in every band, meaning refinement is faithful at all scales. It also fails if $D_b$ tracks F0-RMSE so closely that it adds no information. Either outcome would be reported as a negative result.
References
- Adapting Speech Language Model to Singing Voice Synthesis
- CoMelSinger: Discrete Token-Based Zero-Shot Singing Synthesis With Structured Melody Control and Guidance
- StylePitcher: Generating Style-Following and Expressive Pitch Curves for Versatile Singing Tasks
- Analysis of Hydrogen Behavior in Cr-Mo Steels by Electron Microscopic Autoradiography
- Observation of Hydrogen Trapping Sites at 77 K Exposure in SUS316L Steel by means of Tritium TEM Autoradiography
- Computational resolution limit: a theory towards super-resolution
- Learning Point Spread Function Invertibility Assessment for Image Deconvolution
- Muskits-ESPnet: A Comprehensive Toolkit for Singing Voice Synthesis in New Paradigm
- Opencpop: A High-Quality Open Source Chinese Popular Song Corpus for Singing Voice Synthesis
- GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks