How Much Vocal Tract Physics Can a Recording Actually See?

Speech scientists keep finding physics that the textbook "straight tube" model of the vocal tract leaves out. Soft walls absorb energy. Side cavities such as the sinuses and the narrow chamber just above the vocal folds add resonances and dips. Above about 7 kHz, sound also bounces sideways across the tract. Each finding is usually followed by a call for richer simulators. None of this work asks whether a given recording carries enough statistical evidence to tell the richer model apart from the simple one. We propose to answer that question directly. For each omitted effect, we compute how many frames of speech, at what signal-to-noise ratio and voice pitch, are needed before the effect can be distinguished from measurement noise. We then test those predictions against what inversion and speaker-identification methods actually recover from real audio.

The idea comes from rare-event physics. There, a signal of a given strength over a given background implies an exposure needed for discovery, and an experiment is judged against that limit rather than against the sophistication of its model. Applied to speech, the same reasoning yields a detectability budget for vocal tract physics. That budget would explain why some sub-cavity parameters "wander" when fitted from short windows, would say when a fixed anatomical prior is harmless, and would separate the part of a synthesizer's quality gap that comes from missing physics from the part that is just noise.

Core Claim

We claim that most reported failures of simplified acoustic models fall into two kinds, and that a simple calculation, done before any data is fitted, tells them apart. Some failures are statistically visible: the spectral difference the omitted physics produces exceeds the noise in the log spectrum accumulated over realistic durations. Others are statistically invisible: the difference is real in simulation but lies below the noise floor of speech recordings. This happens especially at high frequencies, where the glottal source rolls off by roughly 6 to 12 dB per octave. It also happens at high pitch, where the transfer function is sampled only at sparse harmonics.

The calculation has two parts. The first is a detectability score for each effect. The second is a Cramér-Rao lower bound, meaning the smallest variance any unbiased estimator can achieve, for each physical parameter. Both are computed while the parameters of the main oral tube are treated as nuisances. Our central testable prediction is that empirical prior sensitivity, model-selection outcomes, and high-band speaker discrimination all switch on where this calculation says they should.

The Detectability Calculation

We build a family of nested forward models. The base model is a one-dimensional tube cascade of the kind used in the differentiable synthesizer of VocalTrax. We extend it in turn with yielding-wall and viscous losses, an epilaryngeal section, a maxillary sinus side branch parameterized as in Kasess and Kreuzer, and a lower-tract resonance pair representing the high-frequency "peak-dip-peak" pattern of Zhang et al.. Where needed, we add a precomputed correction for transverse cross-modes, taken from 3D simulations of the kind used by Takemoto et al.. Precomputing it keeps 3D solvers out of the inner loop.

For an effect that changes the log-magnitude transfer function by $\Delta_{t,k}$ at the $k$-th harmonic in frame $t$, the expected log-likelihood ratio between models is approximately

$$ D(T) = \frac{1}{2}\sum_{t=1}^{T}\sum_{k} \frac{\Delta_{t,k}^{2}}{\sigma_{t,k}^{2}}, $$

Here $T$ is the number of frames and $\sigma_{t,k}^{2}$ is the variance of the log-spectral estimate at that harmonic. That variance is set by the local signal-to-noise ratio: it grows roughly as its inverse at high ratios and saturates when noise dominates. The quadratic form follows from a Gaussian approximation to log-spectral errors. The effect counts as detectable at significance $z$ when $2D(T) \ge z^{2}$. The required number of frames therefore scales as $z^{2}$ divided by the per-frame information. This is the same scaling as the $1/\sqrt{N}$ counting limit that broadens confidence intervals for sparse cosmic-ray events (Schoorlemmer et al.).

The essential refinement is that $\Delta_{t,k}$ is computed after projecting out whatever the main tube parameters can absorb. For parameter estimation, this projection corresponds to the Schur complement of the Fisher information matrix with respect to the nuisance block. This step is what makes the calculation honest. A sinus dip that the oral tube can mimic by shifting a formant contributes no information, however large it looks in simulation. We use the information-geometric treatment in Cramer-Rao Lower Bound and Information Geometry as the method source. The contribution here is the speech-specific noise model: harmonic sampling at the fundamental frequency, the source roll-off, and aspiration noise filling the gaps between harmonics during breathy phonation.

Research Program

The program has three linked analyses that use the same machinery.

Prior sensitivity of the sinus parameters. Kasess and Kreuzer found that estimates of the sinus bandwidth were highly sensitive to the prior, and that broad priors produced physically implausible clusters. Their data were 40-ms nasal windows at 8 kHz from TIMIT. We will compute the Cramér-Rao bound for sinus frequency, bandwidth, and coupling ratio under exactly their conditions. Our prediction is that the bound on bandwidth exceeds the width of their narrower prior (about 440 Hz), while the bound on frequency does not. That would explain why constraining frequency helped but bandwidth stayed prior-driven. The analysis turns their empirical observation into a falsifiable statement. It also predicts how much longer the pooled windows must be before the bandwidth estimate stops depending on the prior.

The high-frequency tension. Zhang et al. argue that speaker identity lives in high-frequency lower-tract features. Takemoto et al. show that one-dimensional models break down in the same band. Neither asks whether the band is above the noise floor in ordinary speech. We will map detectability as a function of frequency band, fundamental frequency, recording signal-to-noise ratio, and duration. The output is a curve of required speech duration for each effect and condition.

Splitting the synthesizer's quality gap. VocalTrax shows a persistent quality gap against human speech. We will split its reconstruction residual into two parts. The first is the part attributable to estimator variance and optimizer randomness, measured by restarts and compared against the noise floor. The second is the part attributable to structural model bias, detected as residual energy that significantly exceeds the predicted noise. The test is a goodness-of-fit statistic computed per band and per frame.

Baselines

The naive reference point is to always fit the simplest one-dimensional model, or alternatively always the richest model, with no detectability reasoning. For the sinus effect, the "simplest" version is the three-tube model of Kasess and Kreuzer, and the reported gap between the sinus model and it is "an average RMS error reduction of 0.2 to 0.3 dB." Our framework should predict where that 0.2–0.3 dB difference is statistically meaningful and where it is within noise. If our predicted significance does not track per-segment fit improvements, the framework fails.

The strongest existing competitor for the model-selection question is Bayesian evidence comparison: the variational free energy that Kasess and Kreuzer's own variational Bayesian scheme already provides. Held-out spectral likelihood under cross-validation is a second strong competitor. Both decide on observed data whether an extra branch is warranted. We will run them on exactly the same segments, with the same forward models and the same prior grid. Our predictions will use no fitting to the test segments. We count as beating these baselines only in a limited sense. Our forward-looking prediction must agree with their per-condition decisions, at a rank correlation of at least the agreement between the two baselines themselves. It must also extrapolate correctly to durations and signal-to-noise conditions not yet fitted, which evidence computed on one dataset cannot do.

For the synthesis question, the strongest existing system is VocalTrax, which reports "48.7% accuracy compared to 11.0% for the baseline" on automated digit classification of AudioMNIST reconstructions (Mo et al.). We do not aim to beat 48.7% by adding physics everywhere. The claim is narrower. Adding only the effects our audit flags as detectable should recover most of whatever accuracy gain adding all effects gives, at lower cost. Adding effects flagged as undetectable should give no significant gain. Every variant gets the same optimizer, restarts, and step budget.

The key ablation removes the nuisance projection, scoring each effect by its raw simulated spectral difference. If raw differences predict empirical outcomes as well as projected ones, the Fisher-information machinery is unnecessary.

Why not just…?

Why not just use a large neural speech model and skip physical modeling? Neural models can reproduce high-frequency speaker cues without telling us which anatomical parameter produced them or whether that parameter is identifiable. On small clinical corpora they also overfit. On the Saarbrücken Voice Database, CNNs reached "around 70% UAR" versus roughly 80–85% for an SVM on engineered features (Automated Detection of Voice Disorder). We still use a neural model, but as an empirical upper probe. A speaker-identification network trained only on the band above 7 kHz measures how much speaker information that band actually carries at each duration. If this probe finds speaker information where our audit predicts none, our noise model is wrong.

Why not just fix the sub-cavities to anatomical means from imaging? Sometimes that is the right answer, and the audit tells you when. Kasess and Kreuzer show that constraining the sinus frequency prior helped while the bandwidth stayed prior-sensitive. Zhang et al. argue that fixing lower-tract cavities to population means erases speaker identity. These claims are compatible only if the cost of fixing a parameter depends on its detectability. We will test this by fixing each parameter at the mean and measuring the loss in fit and in speaker discrimination. The prediction is that the loss is negligible exactly where the Cramér-Rao bound exceeds the population spread of that parameter.

Why not just run Bayesian model comparison on each dataset? This is the real threat. If evidence comparison on observed data already answers every question, our contribution shrinks to a design calculator. We settle it with an extrapolation test. We fit on short windows, predict outcomes at four-times longer pooling and at lower signal-to-noise ratios, and check those predictions against evidence comparison actually run at those settings. Agreement shows that the calculation captures the mechanism. Disagreement means the evidence computation sees structure that our Gaussian log-spectral approximation misses. We would report that outcome as a finding, not hide it.

Proposed Experiments

The first decisive check uses simulation with known ground truth. We generate audio from the richest model, using random anatomy drawn around imaging-based means, a range of fundamental frequencies from 100 to 400 Hz, and added noise at controlled signal-to-noise ratios. We then measure the empirical rate at which likelihood-ratio tests detect each effect. The predicted and observed detection curves must agree within sampling error across conditions. Without this, nothing else stands.

The second check is a replication on TIMIT. We use the 619 nasal-stop segments from dialect region 1 at Kasess and Kreuzer's settings, then pool frames per speaker. Their bandwidth prior sensitivity should appear and then disappear at the duration our bound predicts.

The third check concerns the high band. TIMIT is sampled at 16 kHz and therefore reaches only 8 kHz. We will use recordings with higher native sampling rates, specifically the Saarbrücken sustained vowels and the GTSinger singing corpus (Zhang et al.), after confirming that their rates support the band. We compare the predicted detectability of peak-dip-peak features against the high-band neural speaker probe as functions of duration and pitch. Singing is the natural stress test, because high fundamental frequencies sample the transfer function sparsely. No reported result was found on Lacuna for high-band speaker identification on these corpora, so all such numbers will be new.

The fourth check is the bias-variance split of VocalTrax on TIMIT and AudioMNIST. We ask whether the band-wise residual exceeds the restart-variance floor, and whether adding only the audit-flagged physics closes that excess.