Not All Time Is Equal: Source-Specific Undershoot Curves in Conversational Vowels

When people speak quickly, their vowels drift toward a neutral, mumbled "uh"-like quality. The textbook explanation, called articulatory undershoot, is mechanical. The tongue and jaw have mass, so when a vowel is given less time they fall short of their target positions. If that story is complete, then what matters is how long the vowel lasts, not why it lasts that long. A vowel stretched by slow overall talking should reach its target just as well as one of the same length stretched by the slowdown speakers produce at the end of a phrase. We propose to test this directly on spontaneous speech. We split each vowel's duration into the parts contributed by different sources: overall talking speed, closeness to the end of the utterance, utterance length, stress, and word frequency. We then estimate a separate smooth curve linking each source's contribution to how far the vowel's formant trajectory gets toward its target. Formants are the resonance frequencies that mark vowel quality.

The social stakes follow from this. Sociophonetic work often reports that older or female speakers reduce vowels more or less than others. Recent work shows that such differences can be artifacts of how long speakers' turns are and how fast they talk (Tanner et al.; Gendrot et al.). If the curve from duration to target attainment depends on where the duration comes from, then controlling for "duration" or "rate" is not enough to clean up a social comparison. The question becomes sharper. Do social groups differ only in how much time their vowels get, or does the curve itself bend differently for them?

Core Claim

We claim that duration gained from phrase-final lengthening buys less spectral target attainment than equal duration gained from slower overall rate. We also claim that most of the apparent age and gender differences in vowel reduction in conversational English disappear once the source-specific curves are in the model. The leftover social signal should be a change in the curve's shape, such as a later approach to the target, rather than a constant offset.

The first claim is motivated by a tension in the file. Kocharov et al. found that pre-boundary lengthening of Russian post-tonic vowels (0.29 z-score) produced no significant reduction in vowel omission or replacement. By contrast, Georgiou et al. report that duration tracks vowel space in Greek development, and Gendrot et al. find articulation rate is the strongest predictor of reduction in French. A single undershoot curve cannot produce both patterns. Separate curves for each source of duration can.

What Exists and How This Differs

Three studies supply the method, and each leaves the source question untouched.

Li & Al-Tamimi fit generalized additive mixed models (GAMMs) to 11-point F1 trajectories of Mandarin diphthongs under tone. A GAMM is a regression whose effects are smooth curves learned from data, with speaker-specific random curves. Their segments are time-normalized and cut off from utterance context, and they use 10 male read-speech speakers. McAuliffe & Babel used smoothing-spline ANOVA on Buckeye formant trajectories. They showed that frequency-driven centralization, along with age differences in it, shows up along the trajectory but is missed at the midpoint. They did not condition on rate, utterance length, or boundary position, and they used only 20 female speakers. Tanner et al. model articulation rate with a 5-knot spline on utterance length across 116,020 utterances, but rate is a per-utterance scalar there, with no link to vowel spectra.

Two further studies constrain the design. Yuan et al. show a three-phase relationship between turn length and rate, which forbids linear or binned length terms. Utterance-level normalization shows that final lengthening begins 5 to 11 phones before the boundary, depending on phone class. That finding tells us to measure boundary position as phone count from the end and to model it as a smooth term, not a binary "final" flag.

None of these papers asks whether duration from different sources has different spectral value. That question is the new part of this proposal.

We also use the normative-modeling idea from longitudinal biomedicine, in a limited and specific role. It does not decide which social groups count as different. Its job is to produce per-speaker deviation curves relative to the pooled structural model, so that individual speakers, and not just group means, can be inspected.

Research Program

The primary data are the Buckeye corpus: 40 talkers, about 870,000 hand-corrected phones, balanced for gender and for age group (under 30 versus over 40). Age is therefore a two-level factor. We will not claim to model lifespan curves.

For every stressed and unstressed monophthong longer than a minimum duration, we extract F1 and F2 at 11 equally spaced points and Lobanov-normalize them per speaker. We exclude tokens with tracking failures, using a cross-check against a second formant tracker. Vowels had only 77.8% inter-transcriber agreement in Buckeye, so we restrict to tokens whose label matches the citation form, and we run a sensitivity analysis on the unrestricted set.

Target attainment for a token is the distance from its trajectory's most peripheral point to the speaker's own vowel target. That target is estimated as the centroid of the speaker's longest stressed, phrase-medial tokens of that vowel. We also analyze the full trajectory.

The model has two linked parts. The first decomposes log vowel duration $d_i$ for token $i$:

$$\log d_i = g_R(r_i) + g_B(p_i) + g_L(n_i) + g_S(\text{stress}_i, \text{freq}_i) + u_{\text{spk}} + \epsilon_i .$$

Here $r_i$ is local articulation rate, computed excluding the token's own word. $p_i$ is the number of phones to the utterance end. $n_i$ is utterance length in syllables, split into speaker-mean and within-speaker deviation following Tanner et al. Each $g$ is a smooth function, $u_{\text{spk}}$ is a speaker effect, and $\epsilon_i$ is residual duration.

The second part models attainment $A_i$ as a sum of smooth functions of each source's fitted contribution, $\hat\ell_{R,i}=g_R(r_i)$, $\hat\ell_{B,i}=g_B(p_i)$, and so on:

$$A_i = h_R(\hat\ell_{R,i}) + h_B(\hat\ell_{B,i}) + h_L(\hat\ell_{L,i}) + h_S(\hat\ell_{S,i}) + h_\epsilon(\epsilon_i) + \text{vowel} \times \text{speaker smooths}.$$

All contributions are on the same log-duration scale, so the undershoot hypothesis makes a precise prediction: every $h$ is the same curve. We estimate both parts jointly as a multivariate Bayesian model in brms. Fitting them together carries the uncertainty from the duration decomposition into the attainment curves, which a two-step plug-in would not. The full-trajectory version replaces $A_i$ with the 11-point F1/F2 trajectory and adds tensor-product smooths of normalized time with each source term.

Social effects enter as age-group and gender factor smooths on $h_R$ and $h_B$. These distinguish a constant offset from a change in curve shape. We compare them with the same factors entered in a model that has no structural terms.

Baselines

The naive reference point is a midpoint linear mixed model. It regresses F1 and F2 at 50% of vowel duration on age group, gender, vowel, stress, and word frequency, with speaker and word random intercepts. This is the analysis most sociophonetic claims rest on. It is the right floor because the social conclusions we expect to revise come from it. We fit it to exactly the same token set. Beating it means two things: lower held-out-speaker error, and age and gender coefficients that measurably shrink or change sign once structure is added.

The strongest existing method on this corpus is the trajectory analysis of McAuliffe & Babel, reimplemented as a GAMM over normalized time. It uses their predictors (frequency, neighborhood density, age) and adds random speaker smooths, but no rate, length, or boundary terms. It is the strongest competitor because it already captures dynamic reduction and its age interaction. They report that word frequency significantly predicts F2 but not F1 on 2,564 vowels from 20 female speakers. No reported result was found on Lacuna for held-out prediction error or target attainment on Buckeye, so we produce both numbers for this baseline ourselves. Every model gets the same smoothing basis dimensions, the same speaker folds, and the same budget for tuning the number of basis functions (k). Beating it means lower leave-speaker-out trajectory error, plus a clear attribution of its age effect to structural sources.

The key ablation is a single-undershoot GAMM. It is identical to the full model except that attainment depends on total log duration through one smooth $h(\log d_i)$. Everything else is unchanged: rate, length, and boundary terms are still allowed to affect duration, and only the separate attainment curves are removed. This isolates the proposal's one new ingredient. If the source-specific model does not beat it on held-out speakers, the core claim fails.

A secondary baseline is a scalar replication of Tanner et al.'s length spline on Buckeye articulation rate. Its purpose is to confirm the rate decomposition behaves as expected. Tanner et al. report a length-driven rate increase of 2.1 to 2.4 syllables per second against 0.7 to 0.8 for age.

Why not just…?

Why not just add total vowel duration as a covariate? This is the most serious alternative, and it is exactly our ablation. If phrase-final and rate-driven lengthening produce the same attainment curve, the single-duration model will match ours on held-out speakers, and we will report that classical undershoot survives in spontaneous English. The Russian result of Kocharov et al. is our reason to expect otherwise. However, it measured omission and replacement in read professional speech, not formants, so it does not settle the question.

Why not just measure midpoints? McAuliffe & Babel show that midpoints missed effects that trajectories revealed in this very corpus. Still, for the source-equality question a midpoint attainment measure might be enough. We will run the source decomposition on midpoint and peak-attainment scalars as well as full trajectories, and report whether the conclusion depends on that choice.

Why not just add an age spline to Tanner et al.'s rate model? That is a cheap and worthwhile analysis, but it answers a different question: rate, not spectral reduction. It also needs continuous age, which Buckeye lacks. It cannot resolve whether time from different sources has different articulatory value.

Why not residualize acoustics on structure first, then regress on demographics (a normative two-step)? When structural and social variables are correlated, as with turn length and age, residualizing first and regressing second biases the social effects. For that reason we fit everything jointly. The normative deviation curves are used only for describing individual speakers.

Proposed Experiments

The decisive check is a posterior difference smooth $h_B - h_R$ over the overlapping range of log-duration contributions. The source-specific claim predicts that this difference is reliably below zero: boundary time buys less attainment. Undershoot predicts a difference near zero. We pair this with a leave-speaker-out comparison of the full model against the single-undershoot ablation, measured as root-mean-square error on normalized F1/F2 trajectories. We commit in advance to treating any improvement smaller than the spread across folds as a null result.

The social check compares age and gender effects on attainment in three models: the naive midpoint model, the McAuliffe-style baseline, and the full model. We report how much of each effect is absorbed and whether what remains appears as a change in curve shape.

For replication, we extract formants from CORPRES. It provides hand-labeled phrase boundaries and prominence across 1.1 million sounds, which lets us add prominence as a separate source. That also connects to the dispute between sonority expansion and hyperarticulation (Mo et al. find F1 rising with prominence in 9 of 10 monophthongs). A positive result in both spontaneous English and read Russian would show that "how much time" and "why that time" have to be modeled separately before any social difference in vowel quality is interpreted.