A vocal tract model can recover the shape of a speaker's airway from the sound they produce. The model treats the airway as a chain of short acoustic tubes of varying width. A recording, however, is never the sound at the lips. It is that sound after it has passed through a microphone, a room and often a codec, and each of these colors the spectrum in ways the tube model cannot tell apart from anatomy. A gentle high-frequency roll-off from a cheap microphone looks like lossy, soft vocal tract walls or a breathy glottal source. A resonance in the room looks like a formant. Observational cosmology spent two decades on the same problem. There, telescope and atmosphere artifacts imitate new physics, and the field's answer was not only better correction. It also built null tests: statistics that physics says must be zero, so that any nonzero value reveals a contaminant nobody thought to model. We propose to carry that discipline into acoustic-to-articulatory inversion.
The key observation is that voiced speech samples the vocal tract's frequency response only at multiples of the fundamental frequency (the pitch, $F_0$). A sinusoidal, or harmonic, representation measures the amplitude of each of these harmonics. As pitch and articulation change from frame to frame, the same absolute frequency is probed under different anatomical states and by different harmonic numbers. The recording channel is the only one of the three factors that stays fixed in absolute frequency across all of these conditions. This plays the role that disjoint redshift slices play in galaxy surveys. The null test is residual correlation that is shared across frames physics says should be independent.
Core Claim
Harmonic amplitudes extracted from recorded speech can be written, in log magnitude, as
$$\log A_k(t) = S\big(k;\theta_g(t)\big) + \log\big|H\big(kF_0(t);\,\mathcal{A}(t)\big)\big| + C\big(kF_0(t)\big) + \epsilon_k(t),$$
where $A_k(t)$ is the amplitude of harmonic $k$ at frame $t$. $S$ is the glottal source spectrum, parameterized per frame by $\theta_g(t)$ and indexed naturally by harmonic number. $H$ is the lips-to-glottis transfer function of a non-uniform transmission-line tract with area function $\mathcal{A}(t)$. $C$ is a session-constant channel response in absolute frequency. $\epsilon$ is measurement noise.
The three terms depend on different coordinates: harmonic index, articulation and absolute frequency. Because of this, they are separable given enough diversity in pitch and articulation. Any leftover channel error produces a specific signature: residuals correlated at a fixed absolute frequency across frames whose articulation and pitch differ.
We claim three things. First, jointly inferring $C$ with the tract and source removes a bias in recovered area functions and formant bandwidths that per-utterance normalization cannot remove. Second, a cross-condition residual statistic gives a calibrated detector for channel contamination the model does not include. Third, pruning the flagged frequency bands, as survey analyses prune contaminated multipoles, trades a known amount of precision for a large reduction in bias.
Relation to Prior Work
The tract model is standard and we reuse it. Mermelstein built the cascade of tube sections from articulatory outlines with a pulse-plus-noise source. Laprie et al. derive centerlines and area functions from cine X-ray for copy synthesis. Both assume the recorded audio equals the radiated audio, and neither asks what the recording chain does to the fit.
Source–tract coupling supplies necessary physics rather than motivation. Kaburagi shows the epilaryngeal input-impedance peak near 4.1 kHz reshapes glottal flow. Uezu and Kaburagi show glottal termination shifts formants. Motoki shows wall impedance changes bandwidths. Together these mean spectral tilt and bandwidth are physically meaningful quantities, exactly the ones a channel roll-off contaminates. That is why a naive fit is biased and not merely noisy.
On the sinusoidal side, we reuse the harmonic machinery. Asgari et al. provide maximum-likelihood harmonic estimation that is robust in noise. Nørholm, Gerkmann et al. give least-squares phase anchoring that we use to keep harmonic tracks coherent. Parametric and neural source-filter vocoders factor speech into source and filter but never separate a channel term. These include the ARX-LF vocoder of Agiomyrgiannakis and Rosec and HN-uSFGAN, whose residual-spectra loss keeps pitch out of the filter. Any channel coloring is simply absorbed into the all-pole or neural filter. That is harmless for resynthesis and harmful for inversion.
From cosmology we take two methods. The first is the null cross-correlation diagnostic and iterative pruning of Agarwal et al.: cross-power between redshift slices that should be uncorrelated flags unknown contaminants, and bins are removed until parameters stabilize. The second is forward simulation-based inference from SimBIG, which puts observational realism into mocks and learns the posterior with a normalizing flow. SimBIG's own note that finite simulation budgets produce conservative posteriors constrains our design. It is the reason we pair amortized inference with an explicit, low-dimensional channel parameterization rather than relying on augmentation alone.
Research Program
The first component is a forward simulator. It uses cascaded two-port transmission-line tracts with lossy yielding walls, an optional nasal branch in the style of Niu et al., and a glottal source coupled to the tract's input impedance, following Kaburagi's frequency-domain formulation. The simulator emits harmonic amplitude tracks.
The second component is a channel prior. It covers smooth microphone responses as low-order spline curves in log frequency, room resonances as a small set of sharp peaks and notches, and band-limiting codecs. The channel is shared across a whole session.
The third component is joint inference. Per-frame source and tract parameters and the per-session channel are estimated together, both by direct optimization of the equation above and by a normalizing-flow posterior trained on simulated sessions.
The fourth component is the null statistic. After fitting, we group frames into bins of articulation (clustered area functions) and pitch. For each absolute-frequency band we compute the residual covariance between bins that differ in both. Under a correct model this covariance is zero up to noise. We calibrate its distribution on channel-free simulations, flag bands exceeding $n\sigma$, and either refit with an enlarged channel model or prune those bands.
The analysis question is identifiability. We will measure how recovery error in $C$ and $\mathcal{A}$ scales with the spread of $F_0$ and articulation in a session. The motivating analogy is that more independent slices give tighter separation of the common contaminant. We expect channel error to fall roughly as the inverse square root of the number of effectively independent pitch–articulation bins, since it is an average over them. Deviations from that scaling would reveal degeneracies with wall loss or source tilt.
Baselines
The naive reference fits the same transmission-line and coupled-source model to harmonic amplitudes with no channel term at all. This is how copy-synthesis pipelines in the style of Laprie et al. are typically run. It is the do-nothing point, and it sets the size of the bias we claim to remove.
The first normalization baseline subtracts the long-term average log spectrum or cepstral mean per session before inversion, then fits the same model. This is the standard engineering fix. We run it on the identical harmonic tracks with the identical optimizer and parameter budget.
The strongest competitor is a channel-augmented amortized inverter. It is a normalizing flow, trained exactly like our flow on the same simulator and the same number of simulated sessions, but with random channels applied as data augmentation and not inferred. This is the SimBIG recipe applied directly, and it is the approach most likely to match ours. Both flows get the same architecture, simulation count and tuning trials.
The ablations isolate the key ingredient. One is joint inference restricted to single-pitch sessions, which removes the $F_0$ diversity that breaks source–channel degeneracy. Another is joint inference without source–tract coupling. The third is the joint model without null-test pruning. An oracle that is given the true channel bounds achievable performance.
Beating a baseline means lower area-function error and formant-bandwidth error on held-out sessions with channels outside the training prior, together with posterior coverage close to nominal. We found no reported result on Lacuna for area-function recovery under recording-channel distortion. The closest numbers in the file, SimBIG's roughly 27% tighter $\sigma_8$ from forward modeling, concern cosmology and are not comparable, so all speech numbers here will be established fresh.
Why not just…?
Why not just normalize the long-term average spectrum? Subtracting the session mean removes the channel and also removes the speaker's average tract and source tilt, which is part of what we want to measure. It cannot remove anything that varies with articulation. For a speaker with a characteristic habitual vowel space, it biases the absolute area function by construction. Our first experiment quantifies this directly by comparing absolute and relative area-function errors. If normalization matches joint inference on absolute error, our first claim fails.
Why not just train an amortized inverter with channel augmentation? This is the real threat. Augmentation makes the inverter marginalize over channels. That can give unbiased but wide posteriors when channels lie inside the prior, and silently biased ones when they fall outside it. SimBIG itself reports conservatism from finite simulations. Our advantage should appear in two places: tighter posteriors when a session has wide pitch spread, because the channel is actually estimated rather than averaged out, and detection of out-of-prior channels through the null test. The out-of-prior experiment below settles this. If augmentation matches us there with calibrated coverage, the null test adds nothing and we would report that.
Why not just fit a free equalizer filter jointly, as vocoders implicitly do? An unconstrained filter absorbs the tract's own tilt and bandwidth information, which is the degeneracy vocoders like HN-uSFGAN tolerate because they only need resynthesis. The identifiability argument says the channel is separable only because it is tied to absolute frequency and shared across the session. The single-pitch ablation tests exactly this: without pitch diversity, the joint fit should collapse to the unconstrained-filter behavior.
Proposed Experiments
The first experiment is fully synthetic. Simulated sessions with known area functions, coupled sources and channels are drawn from the prior and from an out-of-prior family (for example unseen room resonances and codec cutoffs). This measures area-function error, formant frequency and bandwidth error, channel error in dB, and posterior coverage. It also measures the scaling of channel error with pitch and articulation spread, and the false-positive and detection rates of the null statistic.
The second experiment is semi-real. We take articulatory recordings with paired imaging, such as the cine X-ray material used for copy synthesis, where reference area functions exist. We then re-record the audio through several physical microphones and rooms, or convolve it with measured impulse responses, so that the true channel is known by construction. The decisive check is agreement of area functions recovered from different re-recordings of the same utterance. Joint inference should make them agree with each other and with the imaging reference, while the naive fit and normalization should drift channel by channel.
The third experiment covers failure modes. Above roughly 4–5 kHz, cross-modes break the one-dimensional model, and this should appear to the null test as a "contaminant." We will check whether the test flags this band on channel-free simulations with a three-dimensional tract. If it does, pruning becomes a way to detect model breakdown and not only channel bias, and we would report it as such.
References
- Computer Simulation of Articulatory Activity in Speech Production
- Articulatory copy synthesis from cine x-ray films
- Frequency-domain representation of source-filter coupling and its effect in the production of voice
- A Simulation Study on the Effect of Glottal Boundary Conditions on Vocal Tract Formants
- Effects of wall impedance on transmission and attenuation of higher-order modes in vocal-tract model
- Improving the accuracy and the robustness of harmonic model for pitch estimation
- Least squares estimate of the initial phases in STFT based speech enhancement
- Towards flexible speech coding for speech synthesis: an LF + modulated noise vocoder
- Unified Source-Filter GAN with Harmonic-plus-Noise Source Excitation Generation
- Characterizing unknown systematics in large scale structure surveys
- A forward modeling approach to analyzing galaxy clustering with S <scp>im</scp> BIG
- Estimation of the acoustic properties of the nasal tract during the production of nasalized vowels