Automatic voice disorder detectors flag hoarse, breathy or unstable voices as possibly pathological. They are built on self-supervised speech models such as HuBERT, which learn general speech features from large amounts of unlabeled audio. These detectors have a documented failure: healthy people over about 55 are routinely flagged as disordered. Normal aging, sometimes called vox senilis, produces the same tremor, breathiness and pitch instability that clinicians associate with disease. A fairness audit on the Saarbruecken Voice Database found exactly this pattern. It also found that a classifier given only age and gender reaches 83% accuracy, so the detectors may largely be learning who is old rather than who is ill (Beyond Global Metrics).
We propose to train the detector so that its internal representation carries no information about age beyond what the clinician-rated severity of the voice already explains. An auxiliary network is trained to guess the speaker's continuous age from the detector's features plus the severity grade. The detector is then trained to defeat that guesser. This is a conditional form of adversarial domain adaptation, a family of methods in which a feature extractor is trained to fool a second network that tries to recover an unwanted attribute. It differs from simply making features "age-blind." Older people really are more often disordered, so blind age removal would also strip out real pathology.
Core Claim
A grade-conditioned, continuous-age adversary applied inside a HuBERT-based voice disorder detector should lower false alarms on healthy older adults and raise detection of mild disorders in young adults. The decisive part of the claim is the mechanism: the gain should come from better ranking within each age band, which score-level fixes cannot produce. The method should also avoid the loss of true pathology signal that an unconditional age adversary would cause.
The conditioning variable is the GRBAS grade, the 0–3 clinical severity rating of overall hoarseness that the Saarbruecken database provides. Concretely, let $z = f(x)$ be the detector's pooled embedding of recording $x$, $g$ its GRBAS grade, $a$ the speaker's age, and $y$ the healthy/disordered label. We train
$$ \min_{f,h}\ \max_{r}\ \ \mathcal{L}_{\text{task}}\big(h(z), y\big) \;-\; \lambda\, \mathcal{L}_{\text{age}}\big(r(z, g), a\big), $$
Here $h$ is the disorder classifier, $r$ is an age regressor that sees both the embedding and the grade, $\mathcal{L}_{\text{age}}$ is a Gaussian negative log-likelihood on age, and $\lambda$ sets the strength of the adversary. The objective is implemented with a gradient reversal layer, which flips the sign of the adversary's gradient as it flows back into $f$. At the optimum, age is approximately independent of $z$ given $g$. Two consequences follow. Age information that tracks real voice quality is kept. Age information that would let the detector raise its score just because the speaker is old is removed.
Relation to Prior Work
Prior work supplies the motivating evidence on both sides. Beyond Global Metrics establishes the false alarms and the demographic shortcut, but it repairs them only with per-group score calibration after training. Automatic age detection in normal and pathological voice shows that age remains recoverable from sustained vowels even in pathological voices (AUC 0.80 overall, 0.84 for women, 0.73 for men). This tells us age leaks into acoustic features independently of disease, which is the information we want to remove.
The method comes from adversarial domain adaptation. The closest existing work is Age-Invariant Speaker Embedding for Diarization of Cognitive Assessments. It places a gradient reversal layer before a binary young/senior classifier on speaker embeddings and halves speaker confusion on elderly Cantonese interviews (2.9% to 1.4%). It also shows that plain multi-task learning, which predicts age alongside the main task, hurts performance. We reuse its gradient-reversal mechanism and its lesson that keeping age information encourages shortcuts. We depart from it in three ways:
- Age is modeled as a continuous variable rather than two bins. Physiological decline does not follow a cut-off, as dependency-level analyses show (Analysing the Performance of ASR for Ageing Voice).
- The adversary is conditioned on severity. In speaker diarization, identity and age are not supposed to correlate, but in pathology detection the label and age genuinely do.
- The target is a diagnostic decision rather than an embedding used for clustering.
One body of work constrains the design. The Invariance Discriminability Tradeoff in Domain Adaptation and FAN-Net document that aggressive invariance collapses class structure. Episodic Training for Domain Generalization notes how sensitive adversarial training is to the weighting schedule. We read these as a prediction about the unconditional adversary: it should damage detection of disordered older speakers. Conditioning is our answer to that risk, and the prediction is directly testable.
ALIGN supplies a practical choice. It applies the adversary to intermediate layers rather than outputs, which helped cross-session neural speech decoding. We therefore compare attaching the adversary to the pooled embedding with attaching it to middle HuBERT layers.
Research Program
Base detector. We follow the setup in Beyond Global Metrics: HuBERT features with a lightweight classifier, trained on the Portuguese AVFAD dataset and the Saarbruecken Voice Database (SVD).
Two evaluation regimes. - Within-corpus: SVD is split by speaker. - Cross-corpus: the model is trained on AVFAD and tested on SVD. This mirrors the cited audit and is where age shortcuts are most likely to break.
Adversary. The age regressor is a two-layer network that receives the embedding concatenated with a one-hot GRBAS grade. AVFAD's severity annotation is not described in the file. If grade is unavailable there, we condition on the binary label for AVFAD training and on grade for SVD.
Strength schedule. $\lambda$ is ramped from zero, following standard gradient-reversal practice, and selected on a validation split by the within-band AUC criterion described below, not by overall accuracy.
Diagnostics beyond accuracy. We train a fresh linear age probe on frozen embeddings, separately within each GRBAS grade. Its age-discrimination AUC is compared with the roughly 0.80 level that Automatic age detection reports from MFCC-based models. Success means within-grade age is near chance while disorder detection is preserved.
Synthetic shortcut stress test. We build training subsets of SVD in which healthy speakers are mostly old and disordered speakers mostly young, and test on the reversed composition. A detector that has learned age will invert its errors on this split. A detector that has learned pathology will not.
Baselines
Reference points. The naive reference is the unmodified HuBERT detector trained with cross-entropy, run on identical splits, features and training budget. Next to it we report the age-and-gender-only random forest from Beyond Global Metrics, which reaches 83% accuracy. This number is a ceiling on what demographics alone explain: a detector whose errors track that random forest's errors is mostly reading demographics. The file reports no HuBERT detector accuracy or AUC on SVD (no reported result found on Lacuna for it), so we will reproduce this number rather than quote it.
Group-wise Platt calibration. This is the strongest existing method on this exact question, from the same paper. It fits a separate affine score correction for each of six age-by-gender cohorts and reaches a normalized cross-entropy of 0.89. We apply it to the naive detector with the same cohorts and held-out calibration data. We also apply it on top of our model, since the two can be combined. Beating it means two things: equal or better normalized cross-entropy, and, crucially, higher AUC within each age band. Per-group affine calibration cannot change ranking inside a group, so a within-band AUC gain is something calibration cannot match by construction.
Unconditional continuous-age adversary. This ablation removes the grade input from the regressor. It isolates conditioning, the key ingredient. We also run a binary young/senior adversary as a direct reimplementation of the Age-Invariant Speaker Embedding design, transplanted to diagnosis. All adversarial variants receive the same $\lambda$ search budget. Our method wins this comparison only if it matches the unconditional versions on false alarms among older healthy speakers while clearly exceeding them on sensitivity for disordered older speakers.
Age-by-label balanced reweighting. Training examples are weighted so that every age band has the same healthy-to-disordered ratio, and the model receives the same number of epochs.
Age-as-input with age-specific thresholds. The detector receives age as an explicit input feature, and decision thresholds are tuned separately per age band.
Why not just…?
Why not just calibrate per age group, as the audit already did? Calibration fixes how confident the scores are, not which recordings rank above which. If a healthy 70-year-old and a mildly disordered 70-year-old receive scores in the wrong order, no per-group affine map can reorder them. Within-band AUC is therefore the measurement that separates the two approaches. If our model does not raise within-band AUC over the naive detector, calibration is the better and cheaper fix, and we will say so.
Why not just rebalance the training data across age and label? This is a real threat. Reweighting removes the marginal correlation between age and label in the training set, and that correlation is a main source of the shortcut. What it cannot remove is age information inside the HuBERT features, which persists regardless of label balance (Automatic age detection). In the cross-corpus setting, that information can still interact with corpus-specific cues. The synthetic stress test and the cross-corpus test settle the question. If reweighting matches the adversary on both, the adversary is unnecessary.
Why not just give the model the speaker's age and set thresholds per age band? This lets the classifier subtract an age offset explicitly, and when age is known at deployment it is a strong competitor. Its weakness is that an explicit age input makes the shortcut easier to use. It also cannot separate breathiness caused by aging from breathiness caused by a lesion when both appear in the same recording. We run it with the same tuning budget. If it matches our model on within-band AUC, we will report that the representation-level intervention adds nothing beyond conditioning the decision.
Why not just strip age with a standard adversary? Older speakers in these corpora are disproportionately disordered, so removing all age information removes some pathology signal along with it. This is the invariance–discriminability collapse described in The Invariance Discriminability Tradeoff. The unconditional ablation measures exactly this loss.
Proposed Experiments and Decisive Checks
All results are reported on SVD speaker-disjoint splits and on AVFAD-to-SVD transfer, broken down by age band (under 30, 30–55, over 55) and by gender. The metrics are:
- false-positive rate on healthy speakers over 55, at a sensitivity fixed on validation data;
- false-negative rate on young speakers with GRBAS grade 1;
- AUC computed within each age band;
- normalized cross-entropy;
- the within-grade linear age-probe AUC.
Each configuration is run with five seeds, and we report confidence intervals on the subgroup metrics, because the smallest cells contain few speakers.
The proposal succeeds only if three results hold together. First, the grade-conditioned adversary raises within-band AUC over both the naive and the calibrated detector. Second, it lowers the over-55 false alarms without the drop in sensitivity for disordered older speakers that the unconditional adversary should show. Third, its errors on the reversed shortcut split stay close to its errors on the matched split, while the naive detector's do not.
If reweighting or age-specific thresholds reproduce all three results, the conclusion is that the shortcut lives in the label distribution rather than in the representation. That would still be a useful negative result for the field.
References
- Beyond Global Metrics: A Fairness Analysis for Interpretable Voice Disorder Detection Systems
- Automatic age detection in normal and pathological voice
- Age-Invariant Speaker Embedding for Diarization of Cognitive Assessments
- Analysing the Performance of Automatic Speech Recognition for Ageing Voice: Does it Correlate with Dependency Level?
- Fan-Net: Fourier-Based Adaptive Normalization for Cross-Domain Stroke Lesion Segmentation
- Episodic Training for Domain Generalization
- ALIGN: Adversarial Learning for Generalizable Speech Neuroprosthesis