Does Post-Focus Compression Reach the Tongue? An Ultrasound Test with Calibrated Nuisance Models

When a speaker stresses one word ("No, MARY bought the lamp"), many languages flatten and shrink the words that follow. Pitch range narrows, loudness drops, and syllables often shorten. This is called post-focus compression. It has been measured in the acoustic signal and at the larynx, the voice box. No one has shown whether the tongue also moves less on those backgrounded words. We propose to measure this with ultrasound tongue imaging. A probe held under the chin films the tongue surface, which shows up as a bright line because sound reflects off the boundary between tongue tissue and oral air. We also vary how many syllables follow the focused word, to test whether any tongue reduction is planned over the whole remaining stretch, the way pitch declination is scaled to sentence length.

The hard part is not recording tongues. It is deciding what counts as a real effect. Three ordinary mechanisms can make post-focal tongue shapes look reduced without any focus-specific motor command. Lower pitch itself shifts the tongue root. Shorter syllables leave less time to reach targets. A probe that slowly slips during long sentences shifts the whole image. Articulatory studies also usually involve one to three speakers, so their effects tend to land at p ≈ 0.01–0.10 and get reported as "trends." We borrow a habit from particle physics: fit the signal together with explicitly modeled background components. We then report the result on a graded scale (hint, evidence, strong evidence), and when nothing is found we state an upper limit in millimetres on how large the effect could be.

Core Claim

The claim has two parts, both of which could fail. The first is about the tongue. Post-focus compression includes supra-laryngeal undershoot: a reduction of tongue displacement on post-focal vowels that remains after we account for three things. These are the tongue shift caused by lower pitch, measured per speaker; the undershoot predicted by shorter duration; and probe drift. The second is about planning. The depth or time course of this residual reduction depends on the length of the post-focal stretch, which we call the runway, and not only on a syllable's distance from the focus. The alternative account is that post-focus compression is laryngeal plus passive, with no residual tongue effect once pitch and duration are accounted for. We expect to report a quantitative upper limit either way, so the study is informative even if the answer is no.

What Exists and How This Differs

Three studies motivate the question by documenting post-focus compression everywhere except the tongue. Focus Acoustics in Mandarin Nominals shows post-focal drops in duration, intensity and pitch using smoothing-spline curves and mixed models, but only acoustically and only in a fixed four-syllable frame. An exploration of voice source correlates of focus uses inverse filtering to show a laxer, breathier voice source post-focally, with one speaker and one sentence. Dynamic Vertical Larynx Actions Under Prosodic Focus tracks larynx height with real-time MRI across five Korean speakers, again in a fixed five-word frame and without tongue contours. Our study differs from all three on two counts. It measures the full midsagittal tongue contour, and it manipulates runway length as an experimental factor.

A direct tension in the literature also motivates the study. Pre-focal rephrasing, focal enhancement and postfocal deaccentuation in French found post-focal pitch compression but no significant post-focal shortening, in one speaker and 64 utterances. Hsu and Xu found significant shortening in Mandarin. If tongue undershoot exists but duration does not change, then reduction is not simply a by-product of running out of time. Our design separates these two possibilities.

Two papers constrain the design. Pitch-driven adjustments in tongue positions: Insights from ultrasound imaging shows, in 23 singers, that tongue root position changes systematically with pitch. Lower post-focal pitch could therefore move the tongue with no focus-specific command. We reuse their functional principal component method as a per-speaker calibration task and do not treat it as an explanation of focus. Articulatory reduction in Mandarin Chinese words shows that repetition shrinks the range of tongue deformation. Post-focal words are typically given (already mentioned), so we must control for givenness separately from focus.

Four papers supply methods. Coarticulation across prosodic domains in Italian: an ultrasound investigation and Smoothing Spline ANOVA for Speech Contour Analysis supply the standard contour-comparison method, which we use as a baseline. Quantification of Transducer Misalignment in Ultrasound Tongue Imaging supplies image-similarity measures for detecting probe slippage. Search for the decay B0→γγ and TENTATIVE DETECTION OF QUASAR FEEDBACK FROM WMAP AND SDSS CROSS-CORRELATION supply the logic of joint signal-plus-background fitting and upper limits. Neutrino-4 anomaly: Oscillations or fluctuations? is a warning: mishandled background degrees of freedom inflated a claimed signal. We therefore count and pre-specify every nuisance parameter.

Research Program

Stimuli. Each item contains a fixed target word with a vowel chosen for clear tongue displacement, such as /i/ or /a/, followed by a continuation of 2, 4 or 6 syllables. A preceding question sets up one of four information-structure conditions:

  • corrective focus on an earlier word, making the target post-focal;
  • broad focus on the whole sentence;
  • focus on the target itself;
  • a givenness control, where the target was mentioned in the context but the focus falls after it.

The target always sits at the same position in the sentence, so time since utterance onset is constant across conditions. Only its focus status and the length of material after it change.

Participants and recording. We record 24 adult English speakers as the primary sample. We add 16 Mandarin speakers as a replication, because in Mandarin lexical tone makes pitch-driven tongue effects both larger and easier to predict. All speakers wear a headset-stabilized probe and are recorded with synchronized audio. Each session interleaves anchor tokens (isolated vowels) for drift estimation. It ends with a calibration block in which speakers sustain the target vowels at five pitch levels, following the Chan et al. paradigm.

Contour extraction. Contours are extracted automatically. The tracker is first validated on the multi-rater "I owe you a yoyo" set from Error analysis of extracted tongue contours from 2d ultrasound images, which reports automated error of 1.76–7.11 mm against manual-rater spread of 0.95–2.11 mm. We will pre-train on adult UTI data from the TaL corpus described in Exploiting Cross-Domain And Cross-Lingual Ultrasound Tongue Imaging Features For Elderly And Dysarthric Speech Recognition. Tracking and all model specification are done blind to condition labels, which are hidden until the analysis code is frozen.

Model. For each post-focal vowel token, we fit tongue contour coordinates along the tongue with a hierarchical functional model:

$$ y_s(u) = \mu_s(u) + \phi(u)\,\mathbb{1}[\text{post-focal}] + g_s(u)\,\Delta F_0 + h(u)\,\Delta d + \delta_s(u,t) + \varepsilon, $$

where the terms are as follows:

  • $y_s(u)$ is speaker $s$'s tongue height at point $u$ along the contour;
  • $\mu_s(u)$ is the speaker's broad-focus contour for that vowel;
  • $\phi(u)$ is the focus-specific signal we care about;
  • $g_s(u)$ is the pitch-to-tongue mapping estimated from that speaker's calibration block and fixed before the focus data are touched, so it acts as a known background shape;
  • $\Delta F_0$ and $\Delta d$ are the token's pitch and duration differences from the broad-focus mean;
  • $h(u)$ is a duration-undershoot term shared across speakers;
  • $\delta_s(u,t)$ is a slow drift term constrained by the anchor tokens and image-similarity scores.

The runway question is tested by letting $\phi$ depend on runway length and on syllable position within the runway. Scanning pointwise across the tongue contour creates a look-elsewhere problem, the same issue as searching many mass bins in physics. We therefore compute global significance by permuting condition labels within speaker. Results are reported as a significance level and a tier: below 3σ is a hint, 3–5σ is evidence, and above 5σ is strong evidence. If no tier is reached, we report the 90% upper limit on the peak magnitude of $\phi$ in millimetres.

Power. Before data collection we simulate power, taking the manual-tracing spread (roughly 1 mm) as the smallest effect worth caring about. The target is a sample able to exclude a 1 mm focus-specific undershoot at 90% confidence if the true effect is zero.

Baselines

The naive reference point is the broad-focus condition. It asks whether post-focal contours differ from the same word, in the same position, under broad focus, compared with plain paired differences in tongue height at each point along the contour. This tells us whether anything happens at all. Any claimed effect must exceed the within-speaker token-to-token variability of this comparison. It must also exceed the manual-rater spread of 0.95–2.11 mm reported by Error analysis of extracted tongue contours from 2d ultrasound images.

The strongest existing analysis is acoustic, and our acoustic baseline replicates it. Following Focus Acoustics in Mandarin Nominals, it runs spline curves on pitch and mixed models on duration and intensity, and we add F1/F2 formant centralization. It is the strongest competitor because it is what the field currently accepts as measuring post-focus compression. If formant centralization tracks every ultrasound effect we find, the tongue data add nothing. No reported result was found on Lacuna for tongue or formant reduction in post-focal position, so there is no prior number to beat. The acoustic reference points available are the focal effects in French (about 20% higher focal pitch and 33.71% focal-syllable lengthening, with no significant post-focal duration change) from Pre-focal rephrasing, focal enhancement and postfocal deaccentuation in French. The acoustic baseline runs on exactly the same tokens with the same speaker random effects. Ultrasound "beats" it if the residual tongue signal $\phi$ predicts focus condition in held-out speakers beyond what formants, pitch and duration jointly predict.

The strongest existing articulatory analysis is standard smoothing-spline comparison with confidence-band non-overlap, as used in Coarticulation across prosodic domains in Italian: an ultrasound investigation. We run it on the same contours with the same smoothing parameters. It is the default a reviewer would demand. Our model beats it in two cases. The first is when the two analyses disagree and the disagreement is explained by an identifiable nuisance, pitch, duration or drift, confirmed on the controls below. The second is when both analyses are null and ours additionally yields a finite upper limit.

Three ablations isolate the key ingredient. Each drops one nuisance term: the calibrated pitch mapping $g_s$, the duration term $h$, or the drift term $\delta_s$. If dropping $g_s$ turns a null into a "significant" tongue effect, that shows naive analyses would misattribute pitch-driven root movement to focus, which is the central methodological claim. The givenness-control condition works as a fourth ablation at the design level. It separates focus-driven reduction from the repetition-driven reduction shown by Articulatory reduction in Mandarin Chinese words.

Why not just…?

Why not just measure formants and skip ultrasound? Formant centralization is cheap, and it may well track part of any tongue reduction. However, formants cannot separate tongue from jaw or lip changes. They also cannot isolate the tongue-root shift that pitch alone produces, which is the confound documented by Pitch-driven adjustments in tongue positions: Insights from ultrasound imaging. Ultrasound has also revealed covert tongue distinctions where acoustics suggested neutralization (Ultrasound Tongue Imaging for Articulatory Measurement). This alternative is a real threat, and the held-out-speaker prediction test settles it. If the tongue residual adds no predictive value over formants, we will say so.

Why not just use standard mixed models with p < 0.05 or confidence-band non-overlap? For detecting a large effect, they would work fine, and we run them. They fail in two specific places. They leave pitch and drift unmodeled, so a pitch-driven or probe-driven shift is reported as focus. They also turn a null into "not significant" rather than into a bound, which is what left the French duration question and the borderline velocity trends in Prosodic position effects and function words in English: a pilot study unresolved. Upper limits could also be obtained from equivalence tests or Bayesian intervals. We do not claim the limit-setting statistics are new. What is new is using calibrated, pre-fixed background shapes, and the ablations show whether they matter.

Why not just use electromagnetic articulography, which gives sub-millimetre points? Glued sensor coils track two or three points on the front and middle of the tongue. The pitch confound acts mainly on the tongue root, where coils cannot be placed, and the coils themselves can perturb articulation (Announcing the electromagnetic articulography (day 1) subset of the mngu0 articulatory corpus). Ultrasound's weakness, probe drift, is exactly what the anchor tokens and the drift term address.

Why not just normalize by duration and drop the runway manipulation? Dividing by duration assumes reduction is spread linearly over the post-focal stretch. Declination does not behave that way in spontaneous speech (Comparing journalistic and spontaneous speech: prosodic and spectral analysis), and phrase boundaries get inserted as utterances lengthen (Japanese prosodic phrasing and intonation Synthesis). More importantly, the runway manipulation is what distinguishes a planned reduction from a reduction that simply decays with distance from the focus.

Decisive Checks

The first check concerns the tongue claim. It requires a residual focus effect $\phi$ at 3σ or more (global, permutation-corrected) after all three nuisance terms, present in both English and Mandarin, and larger than the givenness-control effect. If the combined English–Mandarin analysis instead gives a 90% upper limit below about 1 mm, we conclude that post-focus compression is laryngeal plus passive and report that bound as the result.

The second check concerns the runway claim. We compare a model where reduction depends on runway length and position with one where it depends on distance from focus alone. The comparison uses held-out-speaker likelihood, and the runway claim requires the first model to win.

The third check validates the drift model. Sessions flagged as misaligned by image-similarity measures must show larger fitted drift, echoing the 219 vs. 322 mean-squared-error contrast in Quantification of Transducer Misalignment in Ultrasound Tongue Imaging. Removing those sessions must not change the estimate of $\phi$ beyond its uncertainty. If it does, the drift component is failing and the tongue claim is withdrawn.