Speaker-Specific Vocal Tract Priors for High-Pitch Glottal Inverse Filtering

When a person speaks or sings, the vocal folds produce a buzzing airflow, and the throat and mouth shape that buzz into vowels. Glottal inverse filtering is the attempt to undo the shaping and recover the original airflow from a recording. Clinicians, voice scientists and speech synthesizers all want that airflow, but recovering it breaks down as pitch rises. A high voice samples the vocal tract's frequency response only at a few widely spaced harmonics. Each vibration cycle is also so short that standard methods run out of samples to fit the vocal tract model. As a result, the people whose voices are hardest to analyze are women, children and sopranos.

This note proposes a simple fix that has not been tried. Estimate a speaker's vocal tract filter where it is easy to estimate, from their own lower-pitched speech where harmonics are dense. Then use that filter as a constraint when inverting the same speaker's high-pitched voice. The filter is allowed to move only in the few directions that real anatomy and articulation actually vary. The anatomical evidence comes from inter-speaker MRI data, which serves two roles. It defines which deformations of the vocal tract are plausible, and it supplies a new benchmark in which real human airway shapes, rather than idealized tubes, generate test vowels with a known true airflow.

Core Claim

High-pitch inverse filtering fails mainly because the problem is underdetermined, not because the optimizer is weak. With fewer than ten harmonics and a closed phase under about 1.5 ms, the acoustic signal alone does not pin down the vocal tract. A speaker's own low-pitch recordings plus anatomical knowledge of how a tract can deform can make up the missing information. We expect a prior-regularized version of the current best high-pitch method to keep glottal-parameter error close to its low-pitch level at fundamental frequencies of 300–450 Hz. We expect it to do so on vocal tracts built from real speakers' anatomy, which is where an idealized-tube validation could be misleading.

The method we build on is the quadratic-programming formulation of Glottal inverse filtering based on quadratic programming. That paper estimates the vocal tract, lip radiation and a non-negative airflow jointly in one optimization, and penalizes airflow during the closed phase. It reports flat error up to 450 Hz, but only on synthetic vowels generated from standardized male, female and child airway models. The quasi closed phase method of Quasi closed phase analysis for glottal inverse filtering uses a weighted fit that de-emphasizes samples around the moment of vocal fold closure. It shares the same idealized evaluation and depends on an external detector of closure instants. Neither method uses any speaker-specific information about the filter. That is the ingredient we add.

Concretely, let $s$ be the speech frame, $g$ the unknown airflow, $a$ the vocal tract filter coefficients, and $W$ the sample weighting. Let $a_0$ be the filter estimated from the same speaker's low-pitch speech for the same vowel, $B$ a small basis of anatomically plausible deformations, and $z$ their coefficients. We solve

$$\min_{g \ge 0,\, z}\; \| W(s - H(a_0 + Bz)\,g) \|_2^2 + \lambda \|g_{\text{closed}}\|_1 + \mu \|z\|_2^2 ,$$

where $H(\cdot)$ is the filtering operator, $g_{\text{closed}}$ is the airflow inside the estimated closed phase, and $\lambda, \mu$ are tuning weights. The filter is never free. It is the speaker's own low-pitch tract plus a few learned deformation modes, such as tract shortening from a raised larynx or wider jaw opening. This shrinks the number of filter unknowns from roughly the order of linear prediction to a handful. That reduction is what should let a 1 ms closed phase be enough. The objective is not jointly convex in $g$ and $z$, so we alternate between solving for $g$, which is the original quadratic program, and solving for $z$, which is a small least-squares problem.

Research Program

The first piece of work is a benchmark. The USC database (Database of Volumetric and Real-Time Vocal Tract MRI for Speech Science) contains 3D MRI of 17 speakers holding 12 vowels. From these scans we extract an area function for each speaker and vowel, meaning the cross-sectional area of the airway from glottis to lips. We drive a transmission-line tube synthesizer with Liljencrants–Fant glottal pulses, a standard parametric model of the airflow shape, whose parameters are known. We sweep fundamental frequency from 80 to 450 Hz. The result keeps the ground-truth design of the Story-model corpora used by the QCP and QPR papers (63,750 and 71,250 vowels), but replaces their idealized tracts with real, idiosyncratic anatomy. That anatomy includes the palate shapes and asymmetries that motor-equivalence work such as Lateral Channel Formation in Australian English /l/ shows vary across speakers.

The second piece is the deformation basis $B$. People do not hold their tract fixed when pitch rises. Singers raise the larynx and change jaw and pharynx posture across registers, as qualitatively described in Articulation skills of singing based on the biomechanical coordination of throat muscles. A low-pitch prior used rigidly would therefore be wrong in a systematic way. We learn $B$ by principal component analysis on the 17 speakers' area functions after removing each speaker's mean. We add an explicit length-scaling mode. Speakers in the benchmark that are held out from building $B$ test whether these modes generalize. The benchmark synthesis applies articulatory perturbations at high pitch, such as shortening and pharyngeal narrowing drawn from the observed inter-speaker spread. The prior therefore always faces a tract that has moved away from its low-pitch reference.

The third piece is optional and borrowed from Cherenkov telescopes. The current best high-pitch methods weight samples using a detected closure instant, and those detectors fail on breathy or noisy voices. The MAGIC telescope faced a similar problem of picking a brief, faint flash out of background light. Improving the performance of the single-dish Cherenkov telescope MAGIC through the use of signal timing did this by requiring that pixels light up within a narrow time window of one another. Background photons arrive at random times, while shower light arrives together. That timing test let the analysis lower its brightness thresholds, and it halved residual background. The speech analogue is that a glottal closure excites all frequency bands almost at once, while noise and resonance left over from the previous cycle do not. We build the weighting $W$ from coincidence of energy onsets across a filterbank instead of from a separate closure detector. We treat this as a secondary hypothesis with its own ablation. Our main claim does not rest on it.

Baselines

The naive reference point is iterative adaptive inverse filtering and plain closed-phase covariance analysis, run with their standard settings. Prior work shows both degrade sharply above 200 Hz, and any method should clear them easily. The strongest existing method is the quadratic-programming approach of Glottal inverse filtering based on quadratic programming. We also report the quasi closed phase method of Quasi closed phase analysis for glottal inverse filtering and the closure-detector-free factorization of Automatic Glottal Inverse Filtering with Non-Negative Matrix Factorization, whose fixed 5.5 ms window is known to straddle multiple cycles above 200 Hz.

For the quadratic-programming method, no reported numeric result was found on Lacuna. The file states only that its error curves for normalized amplitude quotient, quasi-open quotient, H1–H2 and harmonic richness stay graphically flat from 330 to 450 Hz on the idealized Story corpus. We will therefore reimplement it and report its numbers on both a Story-style corpus and the new MRI-anatomy corpus. All methods get the same frames, the same filter order, and the same closure instants where these are needed. Every tunable weight is chosen by grid search on the same held-out tuning speakers with an equal number of trials.

Three ablations isolate the key ingredient. The first uses a generic prior, a population-average tract for the vowel instead of the speaker's own. If this matches the speaker-specific prior, then speaker specificity is not what helps. The second uses a rigid prior with $B$ removed. This shows whether modeling articulatory change across pitch matters. The third uses an oracle prior built from the true high-pitch area function, which bounds how much any prior could help. Beating the strongest method means a statistically reliable reduction in median absolute error on all four glottal parameters in the 300–450 Hz band across held-out speakers, without a loss below 200 Hz. The coincidence weighting is compared against AME weighting driven by a standard closure detector, on clean audio and on audio with added noise and breathiness.

Why not just…?

Why not just shorten the linear-prediction window to fit inside the closed phase? This has already been shown to fail for a basic reason. Above about 300 Hz the closed phase holds fewer samples than the filter has coefficients, so the fit is ill-conditioned. Both DC-constrained linear prediction for glottal inverse filtering and the QPR paper document this. Our prior attacks the same sample shortage from the other side, by cutting the number of unknowns instead of fighting for more samples.

Why not just use the quadratic-programming method as published, since it already stays flat to 450 Hz? This is the real threat. The flatness was shown only on idealized tubes. A method can do well on smooth synthetic tracts and still fail on real anatomy. If our reimplementation of QPR stays flat on the MRI-anatomy corpus as well, the prior is unnecessary. We would report that, and the benchmark itself would remain the contribution. The comparison of QPR with and without the prior on held-out speakers settles this directly.

Why not just train a neural network end-to-end on synthetic vowels to output the airflow? A network faces the same missing information between widely spaced harmonics, as Speech to chant transformation with the phase vocoder and Glottal source shape parameter estimation using phase minimization variants show for spectral methods. Unconstrained image networks are also known to latch onto speaker anatomy and generalize poorly: Towards Automatic Speech Identification from Vocal Tract Shape Dynamics in Real-time MRI reports 42% top-1 accuracy on unseen speakers. Still, a network could implicitly learn a population prior. We will train one on the same synthetic budget, give it the speaker's low-pitch reference as an extra input to make the comparison fair, and test it on held-out MRI speakers.

Why not just use the MRI tract directly at test time? MRI is not available for ordinary recordings, and the midsagittal real-time scans omit the glottis entirely. The prior is designed to need only ordinary low-pitch audio. MRI is used to build the deformation basis and the benchmark, and appears only as the oracle upper bound.

Proposed Experiments

The decisive experiment uses the MRI-anatomy synthetic corpus. Methods are tuned on three speakers and tested on the remaining fourteen. Error on the four glottal parameters is reported as a function of fundamental frequency. The pattern that supports the claim is this: the speaker-specific prior with deformation modes stays near the oracle in the 300–450 Hz band, the generic and rigid priors fall in between, and plain QPR degrades on real anatomy more than it did on Story tracts. A second check measures how robust the prior is to mismatch. We increase the amount of synthetic articulatory change between the reference and high-pitch vowels and track how error grows. This reveals when the prior turns from helpful to harmful.

A real-voice check uses GTSinger (GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks, GTSinger/GTSinger). It has no airflow ground truth, so the check is indirect. For each singer and vowel we take low notes as the reference. We then measure whether recovered airflow parameters on high notes vary smoothly with pitch rather than jumping erratically, and whether resynthesis from the recovered source and filter matches the recording. We treat these as consistency checks, not accuracy measures.

A final transfer test asks whether cleaner source–filter separation helps downstream. Towards Speech Classification from Acoustic and Vocal Tract data in Real-time MRI reports 87.86% accuracy on 54 vowel–consonant–vowel items by fusing real-time MRI video with raw spectrogram magnitudes, against 80.24% for video alone. We replace its audio input with separate filter-envelope and airflow features under the same splits and training budget. USC speakers talk at modal pitch, so we expect at most a small gain there. To make the test pitch-sensitive, we also evaluate on audio pitch-shifted upward, where raw spectrograms should suffer and separated features should not.