In Mandarin, a low "Tone 3" syllable turns into something that sounds like a rising "Tone 2" when another Tone 3 follows it. This is tone sandhi, a rule by which neighbouring tones change each other. A long-running question is whether the change is complete. If traces of the original tone survive, the change is an "incomplete neutralization." Large corpus studies say yes. Mandarin Tone Sandhi Realization: Evidence from Large Speech Corpora finds that sandhi syllables rise less than true Tone 2 syllables across about 2,300 speakers, at p < 0.01. A careful study of Taiwanese says the opposite kind of thing. Tonal Variation and Word Meaning in Taiwanese shows that once each word's own pitch habits are modelled, knowing whether a token is "sandhi" or "citation" adds nothing for pre-boundary tokens. On that account, apparent incompleteness is word identity in disguise.
These two results can both be true only if the Mandarin effect comes from comparing different sets of words, for example ones that differ in frequency. That possibility is serious, because the Mandarin study itself reports frequency effects on the size of the rise. This note proposes to settle the question with a design borrowed from particle physics' handling of weak signals. We would freeze the analysis pipeline before looking at the target contrast. We would estimate the pipeline's false-positive rate on "placebo" contrasts where the true difference is zero by construction. We would then report the word-controlled residual either as graded evidence or, if it vanishes, as an upper limit on how large it could be. The handle that separates a phonetic account from a lexical one is prosody. The same Tone 3 words occur before Tone 3 partners across prosodic boundaries of different strength, so the two accounts make different predictions about how the residual should vary with boundary strength.
Core Claim
The published p < 0.01 for Mandarin sandhi incompleteness does not yet measure incomplete neutralization. It measures a mix of three things. One is the residual tonal trace. The second is the difference between the words that happen to fall in each class. The third is pipeline systematics from forced alignment and pitch tracking. We claim that the residual can be isolated with three ingredients. The first is word-level random effects, as in Chuang et al. The second is a within-speaker ratio of rise magnitudes, which cancels speaker-level baselines such as gender, age and region. The third is an empirically calibrated null built from placebo comparisons. Once isolated, the residual's dependence on the strength of the boundary between the two Tone 3 syllables is diagnostic.
Under a phonetic, gradient account, a weaker coupling between the two syllables should leave more of the original tone behind. That predicts a residual that grows with boundary strength until sandhi stops applying. Under a lexical account, the same word carries the same idiosyncratic contour wherever it appears. That predicts a residual that stays flat across boundary strength once word identity is modelled. Neither prior paper can make this test. The Mandarin corpus work has no word control and no boundary labels. The Taiwanese work has both, but Taiwanese sandhi is a positional tone circle, not the T3-before-T3 rule.
Where This Sits in Prior Work
Three papers set the problem. The Mandarin large-corpus study supplies the claim to be stress-tested and the metric we reuse, the log range of the F0 rise. F0 is fundamental frequency, the acoustic correlate of pitch. That study's demographic results (gender, age 20–29 versus 30–39, Fujian versus Heilongjiang) are exactly the speaker-level systematics our ratio is designed to cancel. Chuang et al. supply the method we inherit, generalized additive models (regression models that fit smooth curves over time) with per-word smooth terms. Their result is the strongest challenge to incompleteness. However, a failure to improve model fit is not a bound, and the authors themselves call word-level meaning "too coarse." Creaky Voice as a Function of Tonal Categories and Prosodic Boundaries motivates a second acoustic channel. It found on COSPRO that creak depends jointly on tone and boundary strength, with Tone 3 sandhi differing from Tone 2 in voice quality. But it could not separate sandhi Tone 3 statistically in its main regression because token counts were too low. Pooling labelled data with a fixed pipeline can address that problem.
Two papers constrain the design. Tonal coarticulation as a cue for upcoming prosodic boundary shows that F0 shape strongly encodes upcoming boundary strength; a random forest predicts it at 92% accuracy. So boundary labels must never be inferred from the F0 contours under test. That would be circular. In unlabelled corpora we therefore infer boundaries only from pause and lengthening cues and from text. A Preliminary Study on Tonal Coarticulation in Continuous Speech finds that tonal coarticulation is absent in 48.9% of tone pairs in running speech, largely at boundaries. This warns that at strong boundaries sandhi may simply not apply. The boundary analysis must therefore model whether sandhi applied separately from how completely it applied.
The statistical machinery comes from the physics papers. From the BABAR Search for the decay $B^0 \to \gamma\gamma$ we borrow two practices: blind analysis and the reporting of an upper limit when no excess appears. From Model-independent $\bar{\nu}_e$ short-baseline oscillations from reactor spectral ratios we borrow the idea of taking ratios so that unknown multiplicative baselines cancel. That paper's own tension with other datasets is a reminder that ratios remove only the systematics they are built to remove. Graded language for intermediate signals, such as the 3.4σ "evidence" in Evidence for a vector charmoniumlike state, supplies the reporting convention. No paper in the file applies any of this to tone or articulation.
Research Program
The quantity of interest is a within-speaker log ratio. For speaker $s$, the rise magnitude $R$ of the first syllable of a sandhi word (T3+T3) is divided by the rise of a matched true-Tone-2 word (T2+T3). The two words are matched on first-syllable segments, frequency band and position:
$$\rho_{s} = \log \frac{R_{s}(\text{sandhi T3})}{R_{s}(\text{matched T2})}.$$
Any speaker-level multiplicative factor in pitch range cancels in $\rho_s$. A mixed model then estimates the population residual $\delta = E[\rho]$. It includes random intercepts for each word and each speaker, and an interaction with boundary strength $b$ between the two syllables. A value of $\delta = 0$ means complete neutralization in rise magnitude. The same structure is applied to F0 contour shape, using by-word smooths following Chuang et al., and to creak probability and spectral tilt measured as in Kuang (2017).
The calibration step is the core methodological contribution. Before unblinding the sandhi-versus-T2 contrast, we construct many placebo contrasts with the same structure but a true difference of zero. In each, lexical T2+T3 words are randomly split into two pseudo-classes matched to the real sandhi set's frequency and segmental profile, and the identical pipeline is run: alignment, pitch tracking, matching and model. The spread of placebo $\hat\delta$ values gives an empirical null that includes alignment and tracking systematics. This is the "background estimate." Significance for the real contrast is quoted against this calibrated null, not the model's nominal standard error. If the real $\hat\delta$ is consistent with zero, we report a 95% upper limit on $|\delta|$. That limit is expressed as a fraction of the canonical T2-versus-T3 rise contrast measured in the same data. This makes "how incomplete could it still be" interpretable without importing an outside threshold.
All matching rules, smoothing choices, the boundary coding and the pre-registered evidence thresholds will be frozen on one corpus before the others are unblinded. The thresholds are a calibrated-z of 3 or more for "evidence" and 5 or more for a claim robust across corpora.
Baselines
The naive reference point is the published analysis: an unmatched comparison of sandhi T3 against all lexical T2 using the log-range rise. We rerun it through our pipeline on the same three ASR corpora (STCMDS, aidatatang_200zh, MAGICDATA) so that any difference from the published result is attributable to analysis rather than data. The reported figure we must reproduce is "p < 0.01 in most comparisons" from the Mandarin large-corpus paper. That paper's summary gives no effect size. No reported result was found on Lacuna for a word-controlled effect-size estimate of sandhi T3 versus T2 rise in Mandarin.
The strongest existing method is the Chuang et al. approach transplanted to Mandarin: a generalized additive model with by-word factor smooths, with a tone-type term added and compared by model fit. It is the strongest competitor because it is exactly the analysis that dissolved incompleteness in Taiwanese. Their summary in the file reports no numerical fit statistics, so the comparison point is qualitative: whether tone type improves fit beyond by-word smooths. It will be run on identical tokens, with the same smoothing basis and the same tuning budget as our model. The proposal beats it in one of two ways. It may detect a residual that the fit-comparison misses. Or it may convert the baseline's "no improvement" into a calibrated upper limit with a stated size. Ideally it does both, in different boundary conditions.
Three ablations isolate each ingredient. Removing the ratio and modelling raw rise with speaker random effects tests whether speaker-level cancellation matters. Removing the placebo calibration and using nominal standard errors tests how much pipeline systematics inflate the published significance. Removing word random effects while keeping everything else tests how much of the effect is word composition.
Why not just…?
Why not just fit the Chuang et al. by-word model on Mandarin and be done? This is the main threat, and it is included as the strongest baseline. Its weakness is inferential, not modelling. "Tone type does not improve fit" leaves open whether the residual is zero or merely below that model's detection power. With only 4,294 tokens over 41 words in the Taiwanese study, the latter is plausible. If our calibrated limit and the by-word fit comparison always agree, the honest conclusion is that the by-word model was sufficient and our added value is only the quoted bound. The boundary-strength test is the part the by-word model cannot answer by itself.
Why not just run an equivalence test or a Bayesian region-of-practical-equivalence analysis? Statistically our upper limit is close kin to these, and we will report a two-one-sided-tests result alongside it. The difference lies in what goes into the error bar. Standard equivalence tests trust the model's nominal standard error. Our concern is that forced alignment and pitch tracking on ASR corpora introduce systematic shifts that nominal errors ignore. The placebo-null ablation directly measures whether this matters. If placebo $\hat\delta$ values scatter no more than the nominal standard errors predict, the calibration is unnecessary and a plain equivalence test would have sufficed. We will say so.
Why not just record articulatory data (EMA or real-time MRI) on sandhi? Articulatory tracking would show whether laryngeal or tongue gestures retain a Tone 3 trace that acoustics hide. Interaction Between Lexical Tone and Intonation: An EMA Study shows such data can separate production models that F0 cannot. But the cohorts in the file are 3 to 10 speakers. Prosodic position effects and function words in English: a pilot study is a case where velocity trends were "borderline or not met," and a sandhi residual against word-level variance would be even smaller. Such data would be unable to separate word effects from tone effects. Settling the acoustic residual first tells a future articulatory study the effect size it must detect. That is the point of an upper limit.
Proposed Experiments
The first experiment replicates and decomposes the Mandarin result on the three ASR corpora. In stages, we add matching, word random effects, the speaker ratio and placebo calibration. We report how the calibrated significance changes at each stage. The decisive outcome is whether the residual survives at calibrated-z of 3 or more after word control.
The second experiment is the boundary test on COSPRO, which provides hand-annotated break indices B1–B5 over 347,196 tokens from six speakers. We take Tone 3 words followed across a boundary by another Tone 3 syllable. Two quantities are modelled: the probability that sandhi applied, and, given that it applied, $\delta$ as a function of break index. A residual that rises with boundary strength after word control supports gradient phonetic incompleteness. A flat residual that is consistent with zero supports the lexical account and yields an upper limit. As a secondary output, the pooled labelled data should finally allow creak and spectral tilt on sandhi T3 to be separated from canonical T3, which Kuang (2017) could not do. The ASR corpora lack break labels. There, boundary strength is coded from pauses and syllable lengthening only, never from F0.
The third experiment is a cross-language check. We apply the same calibrated-limit procedure to the Taiwanese Southern Min spontaneous corpus used by Chuang et al. If their null holds, it becomes a numerical bound. If the calibrated analysis shows a residual where their fit comparison did not, that would directly contradict their conclusion and would need to be reported as such.
The whole program is falsified in a useful way if placebo contrasts in the ASR corpora routinely reach the published p < 0.01 level. That would show the original incompleteness evidence is within the pipeline's own background.
References
- Mandarin Tone Sandhi Realization: Evidence from Large Speech Corpora
- Tonal Variation and Word Meaning in Taiwanese
- Creaky Voice as a Function of Tonal Categories and Prosodic Boundaries
- Tonal coarticulation as a cue for upcoming prosodic boundary
- A Preliminary Study on Tonal Coarticulation in Continuous Speech
- Search for the decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- Evidence for a vector charmoniumlike state in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msubsup><mml:mi>D</mml:mi><mml:mi>s</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:msubsup><mml:mi>D</mml:mi><mml:mrow><mml:mi>s</mml:mi><mml:mn>2</mml:mn></mml:mrow><mml:mo>*</mml:mo></mml:msubsup><mml:mo stretchy="false">(</mml:mo><mml:mn>2573</mml:mn><mml:msup><mml:mo stretchy="false">)</mml:mo><mml:mo>−</mml:mo></mml:msup><mml:mo>+</mml:mo><mml:mi mathvariant="normal">c</mml:mi><mml:mo>.</mml:mo><mml:mi mathvariant="normal">c</mml:mi><mml:mo>.</mml:mo></mml:math>
- Interaction Between Lexical Tone and Intonation: An EMA Study
- Prosodic position effects and function words in English: a pilot study