Forbidden Channels for Detecting Vocal Tract Asymmetry

Most fast models of the vocal tract assume a mirror symmetry. One-dimensional tube models and the newer "2.5D" solvers both treat the left and right halves of the mouth and throat as identical. That assumption is what makes these models cheap. It is also known to fail for lateral sounds like /l/, for asymmetric side pockets in the lower throat, and above roughly 4.5 kHz. Nobody can currently say how large an asymmetry has to be before a recording could even reveal it.

We propose to borrow a trick from particle physics. To isolate tiny violations of flavor symmetry, physicists measure decays that symmetry strictly forbids, so any observed count is pure signal with no symmetric background to subtract. The acoustic analogue is the difference between two microphones placed mirror-wise beside the lips. If the tract and the voice source are exactly mirror-symmetric, this antisymmetric channel carries no sound at all. Any asymmetry appears there at first order. In an ordinary single-microphone spectrum, the same asymmetry appears only at second order and is confused with ordinary shape changes. The project turns that observation into a solver, a sensitivity law, and an inversion method, and tests it on simulated tracts and 3D-printed ones.

Core Claim

Mirror reflection splits every sound field in the tract into a mirror-even part $p_S$ and a mirror-odd part $p_A$. Let $\varepsilon$ be the amplitude of a left–right geometric perturbation. Any observable that is itself mirror-even, such as the spectrum at a microphone on the midline, must then be an even function of $\varepsilon$. Formant shifts therefore begin at $O(\varepsilon^2)$. The odd channel instead obeys

$$p_A(f) = \varepsilon\, g_A(f) + O(\varepsilon^3),$$

where $g_A(f)$ is a frequency-dependent template that a simplified model can compute for a given perturbation shape. With $N$ independent analysis frames at per-frame noise level $\sigma$, matched detection against the template gives a minimum detectable asymmetry $\varepsilon_{\min} \propto \sigma/\sqrt{N}$. Fitting a midline spectrum gives $\varepsilon_{\min} \propto (\sigma/\sqrt{N})^{1/2}$. Both follow directly from the order at which $\varepsilon$ enters the signal.

The claim is that this change of scaling exponent, from $N^{-1/2}$ to $N^{-1/4}$, is real and measurable. It should hold up to a systematic floor set by how well the two microphones are matched. It also gives the first principled answer to the question "when does the symmetric simplified model suffice?" The answer is: whenever the true asymmetry sits below $\varepsilon_{\min}$ for the recording conditions at hand.

There is also a frequency structure to exploit. Below the first transverse cutoff, odd modes excited at an asymmetric junction are evanescent, meaning they die out within a short distance. They reach the lips only when the asymmetry lies near the lips. Above the cutoff they propagate. The frequency profile of $g_A$ therefore carries information about where along the tract the asymmetry sits. That location information is unavailable to one-dimensional area functions.

Relation to Prior Work

The method lens comes from flavor physics. Improved measurement of the branching ratio of J/psi -> K_S^0 K_L^0 and Observation of the Decay psi(2S) -> K^0_S K^0_L measure an SU(3)-forbidden channel and report a ratio between two resonances, the quantity $Q_h$. Normalizing in that way cancels common systematics. We copy both moves by measuring the odd/even power ratio rather than absolute odd power.

Measurements of the branching fractions of psi(3686)->Sigma^0 Lambda + c.c. takes a dedicated off-resonance sample to model background that interferes with the signal. Our counterpart is a symmetric phantom measured in the same microphone setup. It calibrates how much even-mode sound leaks into the difference channel through microphone mismatch, the systematic that physicists would call "continuum." The analogy is methodological only. No QCD physics transfers.

The acoustic side supplies the solver and frames the gap. Computation of the acoustic characteristics of vocal-tract models with geometrical perturbation computes higher-order modes in cascaded rectangular tubes. It shows that random axis offsets of 8% leave low frequencies stable but shift peaks and create spectral zeros above about 4.5 kHz. It looks only at single-point transfer functions and never separates even from odd responses. We reuse its mode-matching machinery and add that separation. 2.5D Vocal Tract Modeling builds mirror symmetry into its depth profile. It reaches Pearson correlation above 0.8 with 3D simulations up to 12 kHz in 120 seconds, against 0.5 hours for 3D. The authors name the symmetry assumption as the model's main limitation. Our work quantifies when that limitation matters, without replacing the solver.

Estimating the vocal-tract area function from formants using a sensitivity function and least square supplies the deterministic sensitivity-matrix idea, converging in about 13 iterations. We extend its sensitivity matrix with odd-channel rows and attach noise models to it. The unsupervised inversion work, Unsupervised Acoustic-to-Articulatory Inversion with Variable Vocal Tract Anatomy and Unsupervised Inference of Physiologically Meaningful Articulatory Trajectories with VocalTractLab, motivates the project. Both papers show that audio alone leaves inversion non-unique and prone to collapse onto a single output. They need anatomy terms or smoothness priors to cope. An observable that is exactly zero under symmetry removes one entire family of degeneracies.

The statistical side constrains the claims. Branching fractions for psi(2S) -> gamma eta' and gamma eta shows a roughly 2σ signal that ends up as an upper limit. Indications of intermediate-scale anisotropy of cosmic rays... shows how scanning many bins inflates apparent significance. For that reason, every detection we report is corrected for the trials incurred by searching over frequency bins.

Research Program

We will build an even/odd-decomposed version of the cascaded-rectangular-duct mode-matching model. Each junction will be expanded to first order in small asymmetric offsets, yielding $g_A(f)$ analytically for a library of perturbation shapes. These shapes include a lateral offset of a single segment, a one-sided side branch, and a one-sided tongue groove. The model will be checked against 3D finite-element runs on a small subset of geometries.

From the decomposed model we will derive the Fisher information for $\varepsilon$ under a stationary noise model for each observable. That gives predicted $\varepsilon_{\min}$ curves as functions of SNR, frame count, frequency band and asymmetry location. These curves are the "sensitivity limits" of the title.

Finally, we will add odd-channel rows to the least-squares sensitivity inversion. This lets it estimate a symmetric area function together with a small number of asymmetry parameters. We will test whether this resolves cases that are degenerate under midline-only data.

Baselines

The naive reference is "do nothing": assume symmetry and use the 2.5D or 1D model as is. Against it we measure how often the symmetric model's predicted spectrum is statistically distinguishable from the truth. Its known accuracy is r > 0.8 against 3D up to 12 kHz (2.5D Vocal Tract Modeling). For asymmetric geometries specifically, no reported result was found on Lacuna.

The strongest physics-based competitor is single-microphone spectral fitting. Here the perturbed-duct model of Motoki and Matsuzaki is fitted to the midline spectrum, including the above-cutoff band where asymmetry is known to show up. It gets the same noise realizations, the same frame counts, and the same perturbation library as our method. We beat it if our method achieves lower $\varepsilon_{\min}$ at matched 3σ trials-corrected significance, and if the advantage grows with $N$ at the predicted exponent.

The strongest learned competitor is a convolutional classifier trained on two-channel (stereo) spectra to decide symmetric versus asymmetric. This mirrors the roughly 25% background-rejection gain a neural network gave EXO-200 (Search for Neutrinoless Double-beta Decay with the Complete EXO-200 Dataset). It will be trained on simulator output with a tuning budget matched to our template-fitting compute. We will report its detection rate at a fixed false-alarm rate next to ours.

Two ablations isolate the key ingredient. The first replaces the template with an unmatched power detector on the difference signal, which should scale as $N^{-1/4}$. The second drops the symmetric-phantom leakage calibration, which tests whether the systematic floor really comes from microphone mismatch. For existing inversion methods, the reported reference is SP-AM's MCD of 4.63 and average correlation of 0.32 on HPRC (Unsupervised Acoustic-to-Articulatory Inversion with Variable Vocal Tract Anatomy). Those figures measure symmetric articulatory inversion and do not test asymmetry, so they set context rather than a direct target.

Why not just…?

Why not just run full 3D finite-element simulations with Monte Carlo perturbations? At 0.5 hours per 3D run versus 120 seconds for 2.5D (2.5D Vocal Tract Modeling), mapping sensitivity curves over thousands of noise-and-geometry combinations is impractical. Brute-force runs also yield empirical variances without the even/odd structure that explains why the scaling exponents differ. We do use 3D runs, but only as ground truth on a subset of geometries.

Why not just fit asymmetry from an ordinary spectrum, perhaps above 4.5 kHz where it is visible? This is the real competitor, and the experiments are designed to test it. Our symmetry argument says midline effects are even in $\varepsilon$, so their sensitivity scales as $N^{-1/4}$ and the fit trades off against symmetric area changes. If the measured single-microphone curve matches our odd-channel curve on the phantoms, the core claim fails, and we will report that.

Why not just train a neural network on stereo recordings? A network given two channels can in principle find the odd signal for itself. If it matches the template detector, the contribution shrinks to a theory of why it works plus a calibrated sensitivity limit. The network cannot locate the asymmetry along the tract or give trials-corrected significance without additional machinery. The inversion papers also show that unconstrained networks collapse onto one output without physical priors (Unsupervised Inference of Physiologically Meaningful Articulatory Trajectories with VocalTractLab).

Proposed Experiments

Simulation. We take the MRI-derived vowel geometries (/a/, /i/, /u/, /I/, /E/, /o/) used in the 2.5D and perturbation studies. We impose controlled asymmetries spanning a decade of $\varepsilon$ at three positions along the tract: lips, mid-tract and pharynx. We add calibrated noise and estimate $\varepsilon_{\min}$ at 3σ and 5σ, corrected for trials across frequency bins, for each method.

Phantom validation. We 3D-print symmetric and asymmetric versions of three geometries and drive them with a midline source. Recordings use a mirrored pair of microphones and a midline reference. The symmetric prints serve as the leakage calibration sample, in the same role as the off-resonance data in the hyperon measurement.

Location and inversion. We test whether the frequency profile of the odd channel recovers asymmetry location. We also test whether the extended least-squares inversion separates pairs of geometries that are indistinguishable from midline data alone.

Decisive checks. First, the log–log slope of $\varepsilon_{\min}$ versus $N$ should come out near $-1/2$ for the odd-channel template and near $-1/4$ for midline fitting, until it flattens at the phantom-measured leakage floor. Second, the odd/even power ratio should scale as $\varepsilon^2$. Third, below cutoff, mid-tract and pharyngeal asymmetries should be invisible in the odd channel while lip-region asymmetries are not. A flat slope difference, or an $\varepsilon^2$ law that fails in the phantoms, would falsify the proposal. A short pilot on lateral /l/ from a few speakers recorded with paired microphones is a stretch goal only. Human source asymmetry is uncontrolled and would need to be bounded first.

References

  1. Improved measurement of the branching ratio of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>J</mml:mi><mml:mo>/</mml:mo><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>ψ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mrow><mml:msubsup><mml:mrow><mml:mi>K</mml:mi></mml:mrow><mml:mrow><mml:mi>S</mml:mi></mml:mrow><mml:mrow><mml:mn>0</mml:mn></mml:mrow></mml:msubsup></mml:mrow><mml:mrow><mml:msubsup><mml:mrow><mml:mi>K</mml:mi></mml:mrow><mml:mrow><mml:mi>L</mml:mi></mml:mrow><mml:mrow><mml:mn>0</mml:mn></mml:mrow></mml:msubsup></mml:mrow></mml:math>
  2. Observation of the Decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ψ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>2</mml:mn><mml:mi>S</mml:mi><mml:mo stretchy="false">)</mml:mo><mml:mo>→</mml:mo><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:msubsup><mml:mi>K</mml:mi><mml:mi>L</mml:mi><mml:mn>0</mml:mn></mml:msubsup></mml:math>
  3. Measurements of the branching fractions of ψ(3686)→Σ¯0Λ+c.c. and χcJ(J=0,1,2)→ΛΛ¯
  4. Computation of the acoustic characteristics of vocal-tract models with geometrical perturbation
  5. 2.5D Vocal Tract Modeling: Bridging Low-Dimensional Efficiency with 3D Accuracy
  6. Estimating the vocal-tract area function from formants using a sensitivity function and least square
  7. Unsupervised Acoustic-to-Articulatory Inversion with Variable Vocal Tract Anatomy
  8. Unsupervised Inference of Physiologically Meaningful Articulatory Trajectories with VocalTractLab
  9. Branching fractions for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ψ</mml:mi><mml:mn/><mml:mo>(</mml:mo><mml:mn>2</mml:mn><mml:mi>S</mml:mi><mml:mo>)</mml:mo><mml:mn/><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:mrow><mml:msup><mml:mrow><mml:mi>η</mml:mi></mml:mrow><mml:mrow><mml:mo>′</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and γη
  10. INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
  11. Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset