Telling Anatomy from Behavior in Gendered Speech with Fast 3D Vocal Tract Models

When two people produce different vowels or sibilants, the difference can come from two very different sources. One is the body: a longer or wider vocal tract shifts every resonance. The other is behavior: where a speaker chooses to place the tongue, how far they spread the lips, how much they exaggerate vowel contrasts. Sociophonetic studies show that gender identity and social role change speech through behavior. Prepubertal children with indistinguishable anatomy already produce gender-differentiated /s/. Among expectant fathers, those who describe themselves as more feminine have larger vowel spaces. The standard measurement tools cannot say how much of a given acoustic difference is body and how much is behavior. Those tools are vowel-space polygons and per-speaker statistical normalization, applied to formants, the resonance peaks of the vocal tract. We propose to answer that question directly. We would fit a physical 3D acoustic model of the vocal tract to each speaker's recordings and read off two separate sets of parameters: anatomy (tract length and cavity proportions) and articulation (tongue and lip configuration). Each comes with a posterior distribution that says how well the data pin it down.

The obstacle is cost. A 3D acoustic simulation of one vocal tract shape is a large finite-element problem, and Bayesian fitting needs tens of thousands of such simulations per vowel token. Model order reduction removes this obstacle. Here it means compressing a simulation with up to a million unknowns into a few dozen by projecting it onto a basis learned from sample solutions. It has turned slow engineering solvers into real-time ones, but it has never been applied to vocal tract acoustics. We would build a parametric reduced model of 3D vocal tract resonances, use it for anatomy/behavior decomposition on existing gender-and-speech corpora, and release it as a small, durable artifact rather than a fragile mesh-and-solver stack.

Core Claim

A reduced-order model of steady-state 3D vocal tract acoustics, parameterized by separate anatomy and articulation modes, is fast enough to support full Bayesian inversion of formants $F_1$–$F_4$ and sibilant front-cavity spectra on every token of a corpus. The resulting decomposition can answer a question current methods cannot. When a gender-associated acoustic difference appears, such as the femininity–vowel-space correlation in fathers, is it carried by inferred articulation or by inferred anatomy? The answer comes with stated uncertainty about whether the data can tell the two apart at all.

The closest existing work is Unsupervised Acoustic-to-Articulatory Inversion with Variable Vocal Tract Anatomy. It fits per-speaker anatomy parameters jointly with articulation, using the one-dimensional tube synthesizer VocalTractLab inside a neural analysis-by-synthesis loop. Its anatomy variable is a learned speaker embedding that clusters by speaker. It gives no posterior on how anatomy trades off against articulation, it never tests social or gender-identity effects, and its tube model cannot represent the 3D acoustic behavior that governs the higher formants and the sibilant region. Our proposal differs on all three points: explicit physical parameters with posterior uncertainty, a 3D acoustic model made affordable by reduction, and a sociophonetic test question.

Why Each Direction Is Here

Gender identity and speech supplies the scientific question. Acoustic Correlates of Parental Role and Gender Identity in the Speech of Expecting Parents measured vowel-space polygons and DCT sibilant contrast in 29 German expectant parents from the AVARE project. It found that in men, self-ascribed femininity correlates with vowel space size ($r = 0.54$, $p < .05$). Raw vowel space area in Hertz also grows when the vocal tract is shorter, so that finding is exactly the kind of result that needs decomposing. In children, Kinsman and Li report gendered /s/ by ages 4–5, and Shetty et al. report different tongue-shape strategies reaching shared acoustic targets. These papers motivate the work: they establish that behavioral, socially driven differences exist even where anatomy is matched.

How sibilant spectra shape gender perception in prepubertal children acts as a constraint. On the LoKiS corpus it found no difference in center of gravity or skewness of natural /z/ between children assigned male and female at birth. Yet morphing the spectral peak shifted listeners' judgments strongly. Any method we build must therefore be sensitive to cues that aggregate spectral moments miss, and must be able to report a null finding honestly.

Formant analysis supplies the measurements and the geometry. Mappings between vocal tract area functions, vocal tract resonances and speech formants for multiple speakers applied principal component analysis to MRI area functions of five speakers. The first two components carry 60.4% of pooled variance and map to backness and height. That work supplies our articulation modes. It also motivates the project: it reports that Bark and Lobanov normalization leave residual anatomical differences when female and male speakers are pooled. Formant trajectories for acoustic-to-articulatory inversion supplies the formant tracking method (candidate selection on linear-prediction peaks plus Kalman smoothing). It also shows that explicit formant tracks improve inversion, which justifies using formants rather than generic spectral features as the likelihood target.

Computational speedup supplies the method. Three lines of work show that projection-based reduction makes stiff or parametric physics cheap while keeping errors small. MP-LSVT for multi-scale transport reaches 319× wall-clock speedup at under 2% error on a reacting combustor. State and parameter estimation for retinal laser treatment shrinks a 3D diffusion problem to 6–7 states for 1 kHz estimation. MORE-Stress reports 153–504× speedup at under 1% error on parametric structural problems. We reuse their tools: offline sampling, projection, and empirical interpolation of nonlinear or non-affine terms.

Breaking the Kolmogorov Barrier with Nonlinear Model Reduction constrains the design. It shows that linear bases fail for moving waves, because the best achievable approximation error decays only like $1/\sqrt{n}$ in basis size $n$. We therefore do not reduce the time-domain wave equation at all. We reduce the frequency-domain problem at fixed geometry, where no wavefront travels. The remaining difficulty is that resonance peaks move as parameters change, and we address it with frequency-banded local bases.

Infrastructure sustainability enters as a design constraint, not a research thrust. The Computing Research Repository argues that tools outlive grants when they are kept in open, unrendered source formats. CloudCAST shows how sensitive speech data can be served under tiered consent. A reduced model is a few megabytes of basis vectors and small matrices. It can be archived in plain formats and run without a finite-element license or a cluster, which makes it far easier to preserve than the full-order pipeline.

Research Program

The full-order model is the steady-state acoustic (Helmholtz) equation on a 3D vocal tract mesh with a radiation condition at the lips. After discretization it reads

$$\big(K(\mu) - \omega^2 M(\mu) + i\omega C(\mu)\big)\,p(\omega,\mu) = f(\omega),$$

where $p$ is the vector of pressure values at the $N \sim 10^5$–$10^6$ mesh nodes, $\omega$ is angular frequency, and $K$, $M$, $C$ are the stiffness, mass and wall/radiation-loss matrices. The source $f$ is a glottal source for vowels and a dipole at the constriction for sibilants. The geometry parameters $\mu = (\alpha, \beta)$ split into anatomy parameters $\alpha$ and articulation parameters $\beta$. The anatomy parameters are tract length, pharynx-to-oral proportion and a width scale. The articulation parameters are the leading area-function principal components from the MRI work, plus lip aperture and constriction location.

Geometry enters through mesh morphing of a reference tract. As a result, $K$, $M$ and $C$ depend on $\mu$ in a non-affine way. Following the retinal and combustor work, we approximate them by empirical interpolation so that the online cost does not depend on $N$. The reduced model projects onto a basis $V \in \mathbb{R}^{N \times n}$ built from snapshots, with a separate basis for each band (below 1.5 kHz, 1.5–3.5 kHz, 3.5–8 kHz):

$$V^\top\big(K(\mu) - \omega^2 M(\mu) + i\omega C(\mu)\big)V\,\hat p = V^\top f(\omega).$$

The predicted transfer function at the lips yields formants $F_1$–$F_4$ and, for sibilants, the front-cavity spectral peak. The basis is built greedily, adding snapshots where an a posteriori residual-based error estimate is largest, as in certified reduced-basis practice (Ghattassi et al.).

Inversion is Bayesian. Each speaker has one anatomy vector $\alpha_s$ shared across all their tokens, and each token has its own articulation vector $\beta_t$. The likelihood compares predicted and Kalman-tracked formants with a measurement-noise model. The prior on $\alpha$ comes from population MRI statistics, with separate prior families for children and adults. Sampling uses Hamiltonian Monte Carlo through the differentiable reduced model.

The output is a joint posterior over anatomy and articulation. We report how strongly they trade off for each speaker, measured as posterior correlation between tract length and the articulation modes. Any speaker-level acoustic measure, such as vowel space area, can then be split into the part attributable to articulation, computed at the speaker's own articulation but a reference anatomy, and the remainder.

The main build risk is data. Only five MRI speakers, one of them female, are listed in the file, and no child geometry. We will widen the anatomy space by morphing these geometries along scale and proportion directions. We will validate the morphing by checking that predicted formants of held-out real speakers fall inside the model's posterior predictive range. If the morphed space cannot reproduce child formants, we will report that limit rather than extrapolate.

Baselines

The naive reference point is the raw measure used in the AVARE study itself: vowel space polygon area in Hertz from mean $F_1$/$F_2$ of the corner vowels, correlated with GEPAQ-F+ femininity. The reported number to reproduce is $r = 0.54$, $p < .05$ for male speakers, from Weirich and Simpson. We run it on the same tokens and the same formant tracks we use. Our method beats it only in the sense of explaining it: we must say whether the correlation survives once inferred anatomy is fixed, with an interval rather than a point estimate.

The statistical-normalization baseline computes vowel space area after Bark conversion and after a vocal-tract-length scaling estimated from $F_3$/$F_4$. It is the strongest cheap competitor for the anatomy question, because it removes a single scale factor without any physical model. We run it on identical tokens. The reduced model counts as better only if, on synthetic speakers with known anatomy and articulation, it recovers the articulation-driven vowel-space change with lower error than the scale-normalized area. On real data it must disagree with the scale-normalized area in at least some cases where held-out formants favor the reduced model.

The strongest existing physical method is analysis-by-synthesis with a one-dimensional tube model. We use the speaker-level anatomy variant of EmJEm, which reports on HPRC a mel-cepstral distortion of 4.63, log-$F_0$ error of 0.46, and mean articulator correlation of 0.32. The utterance-level variant reached 0.34. To keep the comparison fair, we place a 1D tube transfer-function solver inside exactly our Bayesian inversion: same parameterization, priors, sampler and likelihood. Only the forward physics changes. This is also the key ablation, because it isolates whether the 3D acoustics matter. We beat it if posterior trade-off between anatomy and articulation is measurably smaller with the 3D model and held-out $F_3$/$F_4$ and sibilant-peak prediction is better. Agreement on $F_1$/$F_2$ is expected and not counted as a win.

As a reference for articulator tracking, the supervised deep inversion of Speaker-Independent Acoustic-to-Articulatory Speech Inversion reports a Pearson correlation of 0.784 on combined EMA data. It is the right ceiling for "predict tongue sensors." We do not expect an unsupervised physical method to match it, and we will not claim to. We report it so readers can see the gap.

For speedup, the reference is the full-order finite-element solver at matched accuracy. The neural-surrogate competitor is a multilayer network trained on the same number of full-order solves as the reduced model's snapshot set, with an equal tuning budget. For a vocal tract reduced model specifically, no reported result was found on Lacuna, so speed and error numbers will be our own.

Why not just…?

Why not just use Lobanov normalization, the field's standard fix for speaker size? Lobanov normalization z-scores each speaker's formants. That rescales each speaker's vowel spread to unit variance, which largely erases the vowel space expansion the AVARE study measures. It removes the behavior along with the anatomy. The MRI comparison paper also reports that it leaves residual anatomical differences when sexes are pooled (Story et al.). We still run the gentler scale-only normalization as a baseline, because it is a real competitor. The synthetic-recovery experiment decides whether a physical model adds anything beyond it.

Why not just use the 1D tube synthesizer, which is already fast? This is the most serious threat. For $F_1$ and $F_2$ of vowels, a plane-wave tube model is probably adequate, and we expect it to agree with the 3D model there. The case for 3D rests on two things. First, $F_3$/$F_4$ carry most of the information that separates tract length from articulation. Second, the sibilant front-cavity resonance, the cue shown to drive gender perception in children (Funk et al.), lies in a band where cross-dimensional acoustic modes are not negligible. The matched 1D ablation settles this. If 1D and 3D give the same posteriors and the same held-out errors in all bands, we will conclude that the reduced 3D model is unnecessary for this question and say so.

Why not just train a neural surrogate on finite-element samples? A network trained on finite-element solves could be just as fast online. Our reasons to prefer projection are that the reduced model inherits a residual-based error estimate we can check at every posterior sample, and that it stays physically structured outside the training distribution, which matters for child and morphed geometries. The physics-preserving stability arguments of Huang et al. and Qin et al. point the same way. That is an argument, not evidence. We will train the surrogate with the same number of full-order solves and compare accuracy on held-out geometries and calibration of the resulting posteriors. If the surrogate matches on both, it is an acceptable substitute, and the decomposition result stands either way.

Why not just put the data and code on a public hosting site? Code and reduced bases can go anywhere, and we will post them openly in plain formats. The sociophonetic recordings are another matter. They come linked to identity questionnaires and, for LoKiS, involve children, so they need consent tiers of the kind CloudCAST implements. We will therefore distribute the model openly and run inference on protected audio under the data holders' consent terms, rather than redistribute that audio.

Proposed Experiments

First, model fidelity and speed. On morphed MRI geometries held out from basis construction, we compare reduced-model transfer functions with full-order solves across 0–8 kHz. We report the error in $F_1$–$F_4$ and in the sibilant peak, wall-clock time per frequency sweep, and the basis size needed per band. We will also show where the Kolmogorov-type slow decay appears as resonances move. The model is good enough if formant error stays below typical formant-tracking error and an inversion of one token finishes in seconds rather than days.

Second, identifiability on synthetic speakers. We generate tokens from known $(\alpha, \beta)$ spanning adult and child-like anatomies, add realistic tracking noise, and invert with the 3D reduced model, the 1D ablation and the normalization baselines. We measure recovery error for articulation-driven vowel-space change and calibration of posterior intervals.

Third, physical plausibility on HPRC and MOCHA-TIMIT. We correlate inferred articulation parameters with EMA tongue and lip sensors, against EmJEm's 0.32 and the supervised 0.784 ceiling. The aim is to confirm the parameters track real articulation, not to win on this metric.

The decisive check is a re-analysis of AVARE (with permission) and LoKiS. In AVARE we ask whether the male femininity–vowel-space correlation is carried by inferred articulation once anatomy is fixed. With 15 fathers, the posterior interval will be wide, and we will report it as such. In LoKiS, inferred anatomy should show no difference between children assigned male and female at birth, which is a sanity check given matched prepubertal anatomy. We then test whether inferred /z/ articulation, specifically front-cavity length and constriction location, separates children whose natural spectral moments did not. A difference there would turn the morphing study's perceptual result into an articulatory account. A null result, alongside good model fit, would itself be evidence that the perceptual cue is not carried by these articulatory parameters.