Doctors who study voice disorders can learn a great deal from how the vocal folds open and close in each cycle. One measure is the open quotient, the fraction of each cycle the folds spend apart. Another is the exact moment the folds snap shut. The most direct way to measure these is electroglottography (EGG), which passes a weak current across the neck through electrodes and records how much the folds are touching. EGG has picked up clinically important differences that ordinary microphone features miss. In one small cohort it separated multiple system atrophy from Parkinson's disease using only open quotient and pitch variability (Electroglottography for the assessment of dysphonia in Parkinson's disease and multiple system atrophy). EGG has practical problems, though. The electrodes are hard to place, particularly on female patients, and they are not available in telemedicine or on a phone. The widely used microphone-only voice tools measure only surface quantities such as pitch, formants, jitter and shimmer (VoiceLab: Software for Fully Reproducible Automated Voice Analysis).
This proposal builds a fast, microphone-only estimator of EGG-derived vocal-fold parameters. It is trained and checked against the synchronized EGG recordings that already exist in the Saarbrücken Voice Database but have barely been used (Automated Detection of Voice Disorder in the Saarbrücken Voice Database: Effects of Pathology Subset and Audio Materials). The estimator is judged mainly by one question: would a clinician running a small pilot study, with 20 to 40 patients, reach the same conclusion using the surrogate as they would using real EGG? Low frame-by-frame error alone does not answer that. The estimator ships as a reproducible, provenance-logging extractor that non-programmers can run.
Core Claim
An acoustic surrogate for EGG is clinically useful when it preserves group-level contrasts at pilot sample sizes, meaning the sign and approximate size of differences between patient groups and controls. Signal fidelity is a weaker test. We claim three things. First, a lightweight model that takes epoch-based acoustic features as input and is supervised by synchronized EGG can recover open quotient and glottal closure instants well enough to keep the direction and most of the size of EGG-measured group differences in resampled cohorts of pilot size. Second, it can do this at a small fraction of the runtime of glottal inverse filtering, which estimates the airflow through the folds by mathematically undoing the effect of the vocal tract. Third, its outputs are interpretable physical quantities. As a secondary result, these quantities should be competitive features for detecting voice disorders.
What is new is the combination of three elements: EGG as the supervision target, physical vocal-fold quantities as the output, and pilot-scale conclusion agreement as the main evaluation. None of the papers in the file does any of the three.
Position Relative to Prior Work
Three source-level studies show that the feature set matters and motivate this work.
- Zero-frequency filtering. This method filters speech around 0 Hz to locate the instants when the vocal folds close. Its eight epoch features beat Praat jitter and shimmer for disorder classification, 92.87% versus 66.66% SVM accuracy, and held up under noise (Zero Frequency Filter Based Analysis of Voice Disorders). However, it was never compared against EGG. It also gives closure instants but not opening instants, so it does not yield open quotient directly.
- Glottal inverse filtering with an autoregressive hidden Markov model. Features computed from the estimated glottal flow generalized slightly better than self-supervised speech models on the Saarbrücken test set, with F1 of 75.27% versus 74.21% for a modified contrastive predictive coding model. The authors call the method slow and ask for a faster surrogate (Comparison of GIF- and SSL-based Features in Pathological-voice Detection). That request is exactly what we answer. Our surrogate is trained against EGG ground truth, which their method never was.
- EGG in a small neurodegenerative cohort. The EGG study (Electroglottography for the assessment of dysphonia in Parkinson's disease and multiple system atrophy) supplies the target quantities, open quotient and pitch standard deviation. Its 43 subjects also show both the promise and the fragility of pilot-scale findings.
Two classification systems constrain the design.
- Multi-source pathology detection. MVP fuses self-supervised embeddings of vowels and sentences and reaches 95.8% AUC on the Saarbrücken database (MVP: Multi-source Voice Pathology detection).
- Gender-aware classification. GeHirNet shows that routing by gender before diagnosing improves MCC from 0.9041 to 0.9363 (GeHirNet: A Gender-Aware Hierarchical Model for Voice Pathology Classification).
We do not try to beat these systems on detection AUC. MVP itself states that it cannot explain its decisions to clinicians. GeHirNet shows that gender is a confound, so every analysis we run is stratified by gender.
The Saarbrücken benchmarking study supplies the evaluation protocol: speaker-independent 5-fold cross-validation with explicit pathology subsets. It also gives a warning. Deep networks trained on spectrograms overfit at this data scale, reaching only about 70% UAR versus about 80% for an SVM on openSMILE features. For that reason our surrogate is kept small.
VoiceLab contributes the software stance: automated parameter choice with every setting and output logged. It covers only surface features and was not validated on pathological voices, and we fill that gap.
Research Program
The first step is to build the ground truth. We compute open quotient, closure instants and opening instants from the Saarbrücken EGG channel using the standard derivative-of-EGG approach. We then flag cycles where the EGG is unreliable, for example because of weak contact or irregular phonation. Ground truth is used only from cycles that pass these checks, and how often cycles are excluded is reported for each pathology.
The second step is the surrogate itself. It takes a frame of audio and the zero-frequency-filter signal around each detected closure. A small convolutional network then predicts three things: a correction to the closure instant, the opening instant, and a per-cycle confidence score. Open quotient follows from the predicted instants.
Training is speaker-disjoint and stratified by gender. Pathology categories are held out across folds so we can test generalization to disorders the model never saw. The deliberately small capacity reflects the overfitting evidence above.
The third step is the pilot-scale evaluation protocol. For a pathology group and matched controls, we repeatedly draw cohorts of size $n \in \{10, 20, 40\}$ per arm. For each draw we compute the gender-stratified effect size of open quotient twice, once from EGG and once from the surrogate. We then report two quantities. The first is the sign-agreement rate. The second is the attenuation ratio:
$$\rho(n) = \frac{\mathbb{E}[d_{\text{sur}}(n)]}{\mathbb{E}[d_{\text{EGG}}(n)]}$$
Here $d$ is the standardized group difference and the expectation is taken over resampled cohorts. Measurement noise uncorrelated with group pushes $\rho$ below 1, so $\rho$ directly measures how much statistical power a pilot study loses by switching to the microphone.
The fourth step is packaging. The extractor ships as a batch tool in the VoiceLab style. Each output row records the model version, the input settings and the per-cycle confidence, so results remain reproducible after software updates.
Baselines
Naive reference. The do-nothing reference has two parts. For parameter recovery, it predicts each subject's open quotient as the mean of their gender-by-group class, which sets a floor that any surrogate must clear. For pilot conclusions, it substitutes VoiceLab/Praat surface measures (jitter, shimmer, pitch standard deviation) for EGG quantities. This mirrors what a clinician without EGG does today. We ask whether these surface measures reproduce the sign of EGG group differences any better than chance-level agreement.
Analytic zero-frequency-filter epochs. This baseline uses the method from Zero Frequency Filter Based Analysis of Voice Disorders with no learning. Closure instants come directly from the filter. Open quotient comes from the best fixed rule we can tune on the training folds. It is the strongest cheap competitor. The reported result for this approach is 92.87% four-disorder SVM accuracy on that paper's own small sustained-vowel cohort. The file reports no EGG-agreement figure for it.
Glottal inverse filtering. This baseline uses the autoregressive hidden Markov model method from Comparison of GIF- and SSL-based Features in Pathological-voice Detection. We derive open quotient from the estimated glottal flow. It is the strongest physics-based estimator in the file, with a reported test F1 of 75.27% on Saarbrücken detection. It may be run on a subset of the data if its runtime makes running it everywhere infeasible. Any such subset is matched exactly for all methods.
Self-supervised embedding probe. This baseline places a regression probe on frozen self-supervised embeddings of the kind used by MVP: Multi-source Voice Pathology detection and trains it to predict EGG-derived open quotient. It is the strongest learned competitor; MVP reports 95.8% AUC for detection. We found no reported result on Lacuna for predicting EGG parameters from self-supervised embeddings.
Fairness and the bar to clear. All methods use the same folds, the same EGG reliability mask, gender stratification, and the same hyperparameter-search budget on the training folds only. The surrogate beats the field if its attenuation ratio at $n=20$ is significantly closer to 1 than that of every baseline, across held-out pathologies. On the secondary task, its interpretable features should reach UAR within a few points of the reported benchmark under the Saarbrücken protocol: about 80% for an SVM on openSMILE features on the all-pathologies subset.
Ablations. Two ablations isolate the key ingredients. In the first, the surrogate is trained with zero-frequency-filter inputs but without EGG supervision, using self-consistency only; this isolates the value of EGG. In the second, the zero-frequency-filter input is removed and raw audio is used alone; this isolates the value of the epoch prior. A pooled-gender variant tests the confound reported by GeHirNet.
Why not just…?
Why not just use the self-supervised probe, since those embeddings already detect pathology at 95.8% AUC? This is a real threat. High detection AUC does not imply that the embedding encodes open quotient. The glottal-filtering comparison found that self-supervised features looked better on validation (78.99% F1) but generalized worse to the test set than physical glottal features. If the probe matches our surrogate on attenuation ratio and on generalization to held-out pathologies, we will adopt it as the estimator. The contribution would then reduce to the EGG-supervised evaluation protocol and the tool. The probe baseline is set up specifically to settle this.
Why not just use analytic zero-frequency filtering, which is fast and robust to noise? Zero-frequency filtering finds closures well but has no direct estimate of the opening instant. Open quotient carried the key EGG finding in the neurodegenerative cohort. Our analytic-filter baseline with a tuned opening rule tests whether the learned correction matters. If it does not, the simpler method wins, and we would report that.
Why not just keep using EGG where it matters? The EGG study itself documents placement difficulty and a cohort of only 43. The hardware is also absent from remote settings. A microphone surrogate with a known attenuation ratio tells a clinician how many extra patients a remote pilot needs to match an EGG pilot.
Why not just wrap the existing tools in an LLM agent like CellAgent? CellAgent automates choosing among existing single-cell tools. Our bottleneck is different: no validated tool exists that turns microphone audio into vocal-fold parameters. A deterministic, logged extractor is more reproducible. An agent layer could come later, but it is not part of this proposal's claim.
Decisive Checks
Three results would confirm the proposal.
- On held-out Saarbrücken pathologies, the surrogate keeps open-quotient effect-size sign agreement above that of the surface-feature and analytic baselines at $n=20$, with an attenuation ratio near 1. This holds separately for each gender.
- Closure-instant timing error against EGG stays low under the added white and babble noise used in the zero-frequency-filter study.
- Runtime is orders of magnitude below glottal inverse filtering.
Two results would falsify it.
- An attenuation ratio indistinguishable from that of the self-supervised probe or the analytic filter.
- Collapse on held-out disorder categories.
If either happens, the pilot-scale protocol remains a reusable contribution: it is a benchmark that tells clinicians how far any microphone-only vocal-fold measure can be trusted when the study has only a few dozen patients.
References
- Electroglottography for the assessment of dysphonia in Parkinson's disease and multiple system atrophy
- VoiceLab: Software for Fully Reproducible Automated Voice Analysis
- Automated Detection of Voice Disorder in the Saarbrücken Voice Database: Effects of Pathology Subset and Audio Materials
- Zero Frequency Filter Based Analysis of Voice Disorders
- Comparison of GIF- and SSL-based Features in Pathological-voice Detection
- MVP: Multi-source Voice Pathology detection
- GeHirNet: A Gender-Aware Hierarchical Model for Voice Pathology Classification
- CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis