Does Focus Make the Voice Tenser or Laxer? A Scaling-Collapse Test of Glottal Pulse Shape

When people emphasize a word, the way their vocal folds vibrate changes, and not only their pitch. Studies disagree on the direction of that change. In English, Irish and Taiwan Min, emphasized syllables are produced with a tenser, more "pressed" voice. In Finnish, emphasized vowels come out laxer and breathier, while the unemphasized words become pressed. No model reconciles the two. We propose that part of this contradiction may be a measurement artifact. The standard voice-quality numbers are ratios of a pulse's time segments to the length of one vibration cycle. Emphasis also raises pitch, which shortens that cycle. If the segments of a glottal pulse do not shrink in proportion to the cycle, those ratios drift with pitch alone, and that drift can reverse the apparent sign of the emphasis effect.

The tool we use comes from physics, where scale invariance is tested by data collapse. The method rescales observations taken at different scales by a fitted power law and asks whether they fall onto one universal curve. If they do, any departure from that curve is the real signal. We estimate, per speaker, how the shape of each vocal-fold pulse scales with cycle length in neutral speech. We then measure emphasis as a departure from that collapsed curve. Finally we test whether the "pressed versus lax" split between languages survives. Either outcome is informative. If the split disappears, a decade of seemingly contradictory findings becomes one result. If it survives at matched pitch, the split is a genuine cross-linguistic difference that a communicative account must explain.

Core Claim

The disputed measures are the normalized amplitude quotient (NAQ) used by Laryngeal voice quality in the expression of focus and the open quotient $O_q$ and $R_d$ used in the Liljencrants–Fant (LF) analyses (An exploration of voice source correlates of focus). NAQ summarizes the closing phase of the airflow pulse; LF analyses fit a standard parametric model of that pulse. All of these measures are dimensionless quotients of pulse durations over the period $T$.

The scaling hypothesis is simple. Within a speaker, assume that a pulse time segment $d$ (for example the effective closing duration) scales with period as

$$ d = c\,T^{\beta}, $$

where $c$ reflects phonation type and $\beta$ is a speaker-level scaling exponent. Under this assumption, the quotient $d/T = c\,T^{\beta-1}$ is invariant to pitch only when $\beta = 1$. If $\beta < 1$, raising pitch raises NAQ with no change in tenseness at all. The obvious correction is to use the collapsed quantity $\tilde{q} = d/T^{\hat\beta}$, which isolates the phonation-type term $c$. We claim the emphasis effect should be defined on $\tilde{q}$, not on $d/T$.

The new contribution is the empirical test of whether the sign contradiction is explained by $\hat\beta \neq 1$ interacting with the different pitch excursions of focus in each language. None of the cited production studies estimate this exponent, and none report focus effects on a pitch-collapsed scale. The Finnish authors themselves hint at a pitch link: they explain breathier focus as "speech economy," meaning speakers avoid raising tension and pitch at once. Our account would replace that physiological story with a mechanical one wherever the data allow.

Why Scale-Invariance Methods, and What They Do and Do Not Import

The gravitational literature supplies a method, not a physical analogy for vocal folds. The Event Horizon Telescope analyses established that Sgr A matches the same Kerr metric as black holes eight orders of magnitude different in mass (First Sagittarius A EHT Results. VI). Okino et al. showed parabolic-to-conical jet transitions recurring across engines of very different mass (Collimation of the Relativistic Jet in the Quasar 3C 273). Both are instances of the same move: rescale by the system's natural size, then compare shapes.

The squeezed three-point work supplies the counterpart lesson. When rescaled curves fail to collapse, as with its factor-of-2 variation in the reduced three-point function against the hierarchical ansatz (Using galaxy pairs to investigate the three-point correlation function in the squeezed limit), the failure itself is the finding.

We borrow both halves. We test collapse of pulse shape across pitch, and we treat non-collapse, such as an exponent that differs between emphasized and neutral speech, as evidence of active laryngeal adjustment.

The skeptic's point that speech is bounded by physiological constants is correct. That point is why we fit the scaling only over each speaker's natural modal pitch range, and why we expect breaks at the register edges.

Research Program

We first build a controlled parallel corpus of all-voiced declarative sentences in Finnish, English and German, with ten speakers per language balanced by gender. The materials follow the narrow/broad/unfocused design of the Finnish study and the all-voiced construction of the Irish English sentence. Speakers produce each sentence in the three focus conditions. They also produce a neutral "pitch sweep" block, in which they read the same sentences at instructed low, mid and high pitch without emphasis. This block lets us estimate $\hat\beta$ independently of focus.

Glottal flow is recovered with Iterative Adaptive Inverse Filtering (IAIF), the method that cancels the vocal-tract filter from the pressure wave, as in the Finnish study. We also fit LF parameters. The all-voiced design limits the formant and nasalization problems that the file flags for inverse filtering of continuous speech.

The analysis has three stages. First, for each speaker, we fit $\log d$ against $\log T$ on the sweep block with a mixed model, and check collapse by plotting time-normalized pulses rescaled by $T^{\hat\beta}$. Second, we re-express NAQ, $O_q$ and $R_d$ on the collapsed scale. Third, we estimate focus effects with generalized additive mixed models, the same tool Dynamic Vertical Larynx Actions Under Prosodic Focus used for larynx trajectories. We report the sign of the focus effect per language on the raw and the collapsed scale.

A secondary analysis tests whether $\hat\beta$ itself shifts under focus. Such a shift would be an active change of vibration mode rather than a ride along the neutral scaling curve.

Baselines

The naive reference point is the published practice: raw NAQ, $O_q$ and $R_d$ compared across focus conditions with no pitch handling, as in the Finnish and Irish English studies. Re-running this on our corpus should reproduce the published sign split. If it does not, the corpus fails as a replication and the main test is void, so this is a necessary sanity check. The cited papers report direction and significance but no effect sizes, so there is no reported numerical baseline for raw focus effects; we found no reported result on Lacuna for the magnitude of NAQ or $O_q$ change under focus.

The strongest competitor on the scientific question is a linear pitch covariate. The same mixed model includes $f_0$, or $\log f_0$, as a fixed and by-speaker random predictor of the raw quotient. This is what a careful phonetician would do, and it already partly separates pitch from tenseness. It is run on identical tokens with the same random-effects structure. The collapse method beats it only if one of two things holds. Either $\hat\beta$ is reliably different from 1 and the collapsed quotient fits held-out sweep tokens better (lower cross-validated error), or the focus-effect signs differ between the two corrections.

The key ablation fixes $\beta = 1$, which is exactly the raw quotient, versus using the estimated $\hat\beta$. This isolates whether the power-law exponent, rather than any rescaling, carries the explanation.

For the downstream check, the strongest existing word-prominence detector in the file reaches 77.2% accuracy from prosodic features alone and 81.5% with lexico-syntactic features on the Switchboard prominence subset (Word prominence detection using robust yet simple prosodic features). We do not expect to beat that system outright, since conversational inverse filtering is unreliable. Instead, we add raw versus collapsed voice-quality features to the same random-forest setup with an equal tuning budget. We count a consistent gain of the collapsed over the raw features as success.

Why not just…?

Why not just include $f_0$ as a covariate and be done? This is the most serious threat, and the experiment is designed to settle it rather than argue around it. A linear covariate on the raw quotient assumes the pitch dependence is additive in the quotient's own units. The scaling relation predicts a multiplicative, exponent-governed dependence that differs across speakers. The comparison of held-out fit and focus-effect sign between the covariate model and the collapsed model will show whether the extra structure matters. If both corrections remove the sign split equally, the scientific conclusion still holds, namely that the contradiction is a pitch artifact, and we will report the simpler correction as sufficient.

Why not just use measures that are not normalized by the period, such as cepstral peak prominence or spectral tilt? Cepstral peak prominence (CPP), a measure of how periodic the voice is, does mark prominence for most listeners in the Bonn data (Cue Interaction in the Perception of Prosodic Prominence). But spectral tilt also covaries with pitch, and these measures do not separate pressed from breathy phonation as cleanly as pulse quotients. We nonetheless compute both on the corpus. If tilt and CPP already show a consistent cross-language focus direction, that independently corroborates the collapsed result. If they also split by language, the split is unlikely to be a quotient artifact.

Why not just accept that languages differ communicatively? That may be the answer. The decisive check is a pitch-matched comparison: we take focused and unfocused tokens whose $f_0$ overlaps within each speaker and compare their raw quotients directly. If Finnish focus is still laxer at matched pitch, the artifact account fails and the cross-linguistic difference is real. Perception results showing listeners lower $R_d$ for prominence even at constant pitch (The Role of Voice Quality in the Perception of Prominence in Synthetic Speech; Voice Source Contribution to Prominence Perception: Rd Implementation) indicate that tenseness is a genuine cue in Irish English. They do not show whether it is one in Finnish.

Proposed Experiments and Decisive Checks

The first decisive check is whether collapse occurs at all. Pulse shapes from the neutral sweep, rescaled by $T^{\hat\beta}$, should overlap better than shapes rescaled by $T$. The confidence interval for $\hat\beta$ should exclude 1 for most speakers. If $\hat\beta \approx 1$ everywhere, the artifact hypothesis is refuted early and cheaply.

The second check is the sign test. We report the focus effect on NAQ and $O_q$ per language on raw, covariate-corrected and collapsed scales, alongside the pitch-matched subset described above.

The third check tests whether focus changes $\hat\beta$ itself. Here we compare exponents fitted on focused versus neutral tokens of overlapping pitch. Following the speaker heterogeneity reported for the Bonn speakers, we examine this per speaker rather than only pooled.

Finally, the downstream classifier comparison on Switchboard and on the Bonn ratings tests whether collapsed voice-quality features add predictive value beyond duration and intensity. Duration and intensity dominated the Bonn random-forest ranking, so a null result there would be informative too.