Spacing, Not Just Dose: A Concentration-Quenching Test for Synthetic Disfluency Augmentation

People who build systems for messy conversational speech often train them on clean text that has been deliberately broken. They insert fake "um"s, repeated words, and abandoned restarts, then teach a model to find and remove them. The only setting these pipelines expose is how much noise to inject. Materials scientists who dope crystals with impurity atoms learned long ago that the amount alone does not decide the outcome. Light emission from a doped crystal rises with dopant concentration and then collapses, and the collapse happens because dopant atoms sit too close together, not simply because there are many of them. This note proposes testing whether synthetic disfluency behaves the same way. The central question is whether, at a fixed overall disfluency rate, the distance between injected self-repairs decides whether augmentation helps or hurts a disfluency detector.

If the answer is yes, the usual practice of sweeping a single "noise intensity" knob is searching the wrong space. Augmentation pipelines would need a second control, a minimum or target spacing between interruption points, and the drop in performance at high noise levels that others have reported would turn out to be a clustering effect that spacing can remove. If the answer is no, that is also useful: it would show that the doping analogy, which the literature invokes loosely, adds nothing beyond a dose sweep, and the experiment is designed so that this outcome is clearly visible.

Core Claim

The physical lens comes from three doping studies, and they supply the hypothesis rather than the method. In Co²⁺-doped persistent-luminescence nanoparticles, Crystal lattice site occupation-based optical properties in Co2+-doped ZGGO persistent luminescence nanoparticles attributes concentration quenching to a critical distance between neighboring dopants, below which energy leaks away as heat instead of light. In Nd³⁺ fluorosilicate glass, Near-Infrared Luminescence and Single-Mode Laser Emission From Nd3+ Doped Compound Glass and Glass Microsphere finds quenching at 0.8 mol% because dopants segregate into fluorine-rich domains, so local clustering triggers failure even when the average concentration looks safe. In nitrogen-doped graphene, Oxygen reduction reactions on pure and nitrogen-doped graphene: a first-principles modeling computes an optimum near 4% nitrogen and irreversible poisoning at 50%. The shared lesson is that a local benefit competes with a penalty from interactions between sites, and the penalty depends on how close the sites are.

The transferable, testable claim is this. A synthetic disfluency, meaning a reparandum (the discarded words), an interruption point, and a repair, gives a detector one clean example of the repair structure. Two synthetic repairs placed within a few tokens of each other produce an overlapping, often unparseable structure that the original data does not license. That teaches the model incorrect span boundaries. The GRASS study gives indirect support for such interaction effects in real speech: What the Filler? Both ASR Systems and Humans Struggle More With Other Kinds of Disfluencies Than With Filler Particles finds that words next to an already-mistranscribed word are wrong 45.63% of the time, and that syntactic incompleteness, not filler particles, drives errors for both humans and recognizers.

A simple quenching-style model makes the prediction sharp. Let $\rho$ be the fraction of tokens that begin an injected repair, and let $d_c$ be a critical distance in tokens. Under uniform random placement, the probability that a given repair has a neighbor within $d_c$ tokens is $1-(1-\rho)^{d_c}\approx 1-e^{-\rho d_c}$. The detector's score then behaves roughly as

$$F(\rho) \approx F_0 + a\rho - b\rho\left(1-e^{-\rho d_c}\right),$$

where $F_0$ is the unaugmented score, $a$ is the per-repair benefit, and $b$ is the per-collision harm. The form follows directly from counting colliding pairs under independent placement. The model predicts that random injection produces a peak in performance followed by roll-off. If placements are instead constrained to a spacing floor $d \ge d_c$, the collision term vanishes and the curve should flatten into a plateau rather than roll off. A plateau under spacing control, at densities where random placement rolls off, is the signature that would confirm the claim.

Research Program

The work builds one controllable injector and uses it to map a two-dimensional response surface: injection density by spacing regime. The injector reuses existing generators instead of inventing a new one. A rule-based injector draws on the three disfluency classes in DACL: Disfluency Augmented Curriculum Learning for Fluent Text Generation, which are repeats, interjections, and false starts, scaled by an intensity parameter $N\in[0,10]$. A learned injector uses the Planner-Generator of Planning and Generating Natural and Diverse Disfluent Texts as Augmentation for Disfluency Detection, which separates choosing where a disfluency goes from writing its reparandum. That separation is exactly what makes placement controllable. The only new component is a placement sampler that sits between "where" and "what." It fixes the total number of repairs per sentence and per batch and varies their arrangement. The four arrangements are uniform random, a hard minimum gap of $d$ tokens, deliberately clustered within short windows, and a fourth that reproduces the empirical spacing distribution of real Switchboard repairs.

That fourth regime matters because real disfluencies are not spread uniformly. Speech Disfluencies occur at Higher Perplexities reports that they concentrate where the upcoming text is hard to predict. Natural speech is therefore "clustered" to some degree. The doping lens leaves open whether matching that natural clustering is optimal or whether a spacing floor beyond it helps. Distance is measured both in surface tokens and in dependency-tree hops, so we can see which one carries the effect.

The detector is held fixed across all conditions as a BERT span classifier following Span Classification with Structured Information for Disfluency Detection in Spoken Utterances. That paper classifies candidate spans using start-token, end-token, and length embeddings, and optionally adds a graph network over dependency parses. We use the version without the parse graph, so that parser errors on fragmented speech cannot confound the placement effect. For each cell of the density-by-spacing grid, we pretrain on synthetic data seeded from WMT2017 News Crawl, as the Planner-Generator work did, and then fine-tune on Switchboard. We fit the equation above to the random-placement curve, estimate $d_c$, and then test whether the spacing-floor curve plateaus at the densities where random placement rolls off.

Baselines

The naive reference point is the same span classifier fine-tuned on Switchboard with no synthetic data, run under the same seeds and tuning budget. It tells us whether any augmentation helps at this detector's strength.

The strongest existing detector in the file is the BERT-GCN span classifier of Span Classification with Structured Information for Disfluency Detection in Spoken Utterances, which reports 94.2 F1 on Switchboard, against 94.1 for its span-based BERT without the graph and 88.2 for a token-level BERT tagger. We reproduce the 94.1 configuration as our base and also report the 94.2 variant. Because the gain from the graph is only 0.1 F1, a placement effect should be measured against seed variance rather than against these headline numbers. Beating this baseline means the best spacing-controlled augmentation exceeds unaugmented span-BERT by more than the spread across five seeds, and does so specifically on sentences containing multiple or nested repairs.

The strongest augmentation competitors are the Planner-Generator, which reports 92.2 F1 with 3M and 92.3 F1 with 20M synthetic sentences on a BERT detector, and DisAug from Boosting Disfluency Detection with Large Language Model as Disfluency Generator. DisAug filters LLM-generated disfluent sentences by token-level uncertainty and finds that about 8K filtered sentences beat larger unfiltered sets. No reported Switchboard F1 for DisAug was found on Lacuna. DACL reports 93.0 F1 and 97.1% word precision, but for a generation (cleanup) task with LCS decoding rather than span detection, so it is a reference point and not a direct competitor. We rerun the Planner-Generator injection and an uncertainty filter in the DisAug style on top of the same detector, with the same synthetic budget measured in tokens, the same fine-tuning schedule, and an equal hyperparameter search. The critical ablation holds density, generator, and data volume fixed and changes only placement, uniform random versus spacing floor. This isolates the single ingredient the proposal adds.

Why not just…?

Why not just grid-search the injection rate? A one-dimensional dose sweep will find the peak of the random-placement curve, and if spacing has no effect it is all anyone needs. The proposal does not argue around this; the ablation settles it. If curves at matched density are indistinguishable across placement regimes, the doping lens is empty and we will report it that way. The claim survives only if the spacing-floor curve stays flat at densities where the random curve falls. A dose sweep cannot reveal that, because it never varies arrangement independently of amount.

Why not just copy the natural spacing statistics of Switchboard? This is the most serious threat. Matching the empirical distribution is simple and principled, and it may be optimal: realism, not a hard floor, could be what matters. That is why natural-matched placement is a full experimental arm rather than a strawman. The doping account predicts a specific difference. Natural clustering at high-perplexity regions should help on test sentences with nested repairs, because it shows the model genuine overlap, while artificial uniform clustering should hurt. If natural matching beats the spacing floor everywhere, the takeaway shifts from "space repairs apart" to "the spacing distribution is a first-class augmentation target." That is still a finding no current pipeline acts on.

Why not just rely on curriculum or uncertainty filtering? DACL reports that exposing a model to the messiest data from the start makes it emit irrelevant tokens, and that staging intensity fixes this. DisAug's filter discards noisy samples. Both may already remove clustered examples implicitly. We test this directly by measuring the spacing distribution of the samples DisAug-style filtering keeps and rejects, and by running DACL's 6-stage schedule under random versus spaced placement. If filtering mostly rejects collisions, spacing control becomes a cheaper explanation of why filtering works, and that outcome would be informative too.

Why not just use a bigger pretrained model? J-j-j-just Stutter: Benchmarking Whisper's Performance Disparities on Different Stuttering Patterns shows that Whisper large-v2 reaches 0.489 semantic WER on stuttered clips against 0.234 on fluent ones and hallucinates on 22.6% of sound-repetition clips. Scale alone does not remove the failure on dense repetition. For text detection, we repeat the key ablation with a larger encoder to check whether the placement effect shrinks with capacity.

Proposed Experiments

The primary benchmark is Switchboard (Penn Treebank-3), using the standard splits and word-level disfluency F1. Results are also stratified by how many repairs a test sentence contains and how far apart they are, because the hypothesis concerns that subset specifically.

The first decisive check fits the quenching model to the random-placement density sweep. It asks whether a single $d_c$, in tokens or in tree hops, explains where the roll-off occurs, and whether the fitted value holds across both injectors. The second check asks whether the spacing-floor arm at $d\ge d_c$ removes the roll-off; the confirming signature is a plateau that stays within seed noise of the peak at high densities. The third check compares the natural-matched and spacing-floor arms on sentences with nested repairs, which separates the realism explanation from the collision explanation.

For out-of-domain transfer, models trained on Switchboard are evaluated on DISFL-QA under each placement regime, to see whether spacing control trades in-domain gains for brittleness. A small secondary probe re-uses SEP-28K manual transcripts, which are 3-second clips with labels for repetition subtypes. It asks whether word-repetition errors made by a text-side cleanup model trained under each regime differ, keeping in mind that the short clips limit syntactic context.

The proposal is falsified if the placement effect at matched density is smaller than seed variance on every stratum.