Most speech-based depression detectors learn from questionnaire scores that patients fill in about themselves, and from interviews in which a scripted interviewer decides what to ask next. Both sources leak information that has nothing to do with how a depressed person's body produces voice. Speakers under-report or over-report. Interviewers ask follow-up questions only after a patient discloses distress. A large pretrained speech encoder will absorb all of this.
We propose a way to restrict a speech encoder to the part of its representation that a physical voice model can account for. We simulate voices with a two-mass vocal fold model, a lumped mass-spring description of the folds. We pass the synthetic voices through the encoder and use principal component analysis (PCA) to find the few directions along which the encoder responds to changes in fold stiffness, damping and lung pressure. Real patient speech is then projected onto only those directions. Because synthetic sustained vowels carry no words, no interviewer prompts and no reporting style, this subspace cannot encode those shortcuts by construction.
The test is whether this physics-anchored representation keeps its accuracy when the label changes. We train it on self-report-labelled interviews and evaluate it on psychiatrist-diagnosed patients. Unconstrained encoders and generic acoustic features are expected to lose much more in that transfer.
Core Claim
The claim is a measurable gap, which we call the label-source transfer gap. It is the drop in detection performance when a model trained against self-report targets (PHQ-8 on DAIC-WOZ, BDI-II on AVEC 2013) is evaluated against clinician DSM-5 diagnoses on the Androids Corpus. We claim this gap is smaller for the physics-anchored subspace than for the full encoder, for the encoder's own top principal components, and for standard acoustic feature sets. The comparison is made at matched dimension.
A second, more falsifiable claim concerns interviewer prompts. On DAIC-WOZ participant-only audio, a model confined to the physics subspace should show no advantage on answers to distress-triggered follow-up questions over answers to neutral questions. Unconstrained models should show such an advantage.
If physically grounded features do not shrink the transfer gap, the argument that biomechanics offers an escape from biased labels is wrong, and that is worth knowing too.
Why This Is Not Already Done
The closest work inverts physical models directly from speech. Classification of stressed speech using physical parameters derived from two-mass model fits stiffness, damping and subglottal pressure to F0 and spectral flatness by simplex search. It shows that these parameters beat acoustic features for acute lab stress, but only in 11 speakers and not for depression. Neurophysiological Vocal Source Modeling for Biomarkers of Disease infers laryngeal muscle activations from F0 contours and reports the best median depression prediction on AVEC 2013, yet it trains and scores against the self-reported BDI-II. Deriving Vocal Fold Oscillation Information from Recorded Voice Signals Using Models of Phonation and Speech-Based Parameter Estimation of an Asymmetric Vocal Fold Oscillation Model and its Application in Discriminating Vocal Fold Pathologies invert a body-cover oscillator by adjoint gradients and classify laryngeal lesions. They never touch psychiatric labels.
These papers supply the motivation and the forward models. All of them treat the physical parameters themselves as the features, which inherits the ill-posedness, local minima and noise sensitivity these papers acknowledge. We instead use the physical model only to define a subspace inside a robust learned representation, and we evaluate on the label-source question that none of them pose.
On the dimensionality-reduction side, Effective Dimensionality Reduction for Word Embeddings shows that the top principal directions of embeddings are often uninformative and should be removed. The number to remove (D=7) is chosen by hand. Functional Principal Component Analysis of Vocal Tract Area Functions shows that five functional components explain 98.4% of vocal tract shape variance across MRI vowels. We reuse the second as a compact generator of plausible tract shapes. We replace the first's label-free but physics-blind removal rule with a physics-defined selection of directions.
On the bias side, DAIC-WOZ: On the Validity of Using the Therapist's prompts in Automatic Depression Detection from Clinical Interviews shows that the interviewer's prompts alone reach F1 0.88 on DAIC-WOZ text. It explicitly leaves the audio modality open. The Androids Corpus: A New Publicly Available Benchmark for Speech Based Depression Detection provides clinician-diagnosed labels but only 32-dimensional acoustic baselines. These two papers define our evaluation.
Research Program
Simulation. We build a synthetic corpus from a two-mass vocal fold model with Bernoulli flow and flow separation, following the stress-classification work. To capture acoustic loading from the tract, we couple it to a wave-propagation tract in the style of Voice production model employing an interactive boundary-layer analysis of glottal flow. Tract shapes are sampled from the span of the five Lucero eigenfunctions, and source parameters (both stiffnesses, damping ratio, subglottal pressure, coupling stiffness and left-right asymmetry) are sampled on a grid covering normal-to-pathological ranges. Each configuration yields a short sustained vowel, rendered with randomized microphone responses and room impulse responses.
Finding the source subspace. We pass every synthetic vowel through a pretrained speech encoder and average its block-8 frame vectors. We use block 8 because Self-Supervised Representations in Speech-Based Depression Detection found it carries the most depression-relevant information in WavLM and HuBERT. For each fixed tract shape and channel, we center the embeddings across source parameters. This removes what the tract and recording conditions contribute. PCA on the pooled, centered embeddings gives an orthonormal basis $U_k$ of the $k$ directions along which the encoder tracks vocal fold physics.
A real utterance embedding $x$ is represented by $z = U_k^\top (x - \mu)$, where $\mu$ is the corpus mean. The dimension $k$ is set by the variance explained on synthetic data, never by depression labels.
Grounding the subspace in real speech. Before any depression experiment, we check that the subspace means something on real voices. A regressor trained on synthetic $z$ should predict physical parameters that agree with independent Nelder-Mead inversions of the same real recordings. On FEMH, $z$ alone should separate lesion types at accuracies comparable to the adjoint-inversion results.
Detection. For depression, $z$ is pooled over voiced segments of each speaker and fed to a regularized logistic classifier. Because the stress-classification work found that male and female speakers express stress through different parameters (damping versus subglottal pressure), all analyses are reported stratified by gender.
Baselines
Naive reference. A majority-class predictor, plus a classifier on age and gender alone. The second mirrors the reference used in the neurocomputational source study on mPower and tells us how much any acoustic system adds.
Strongest existing method. The full 768-dimensional WavLM block-8 representation pooled per speaker with the same classifier, following Self-Supervised Representations in Speech-Based Depression Detection. That paper reports a DAIC-WOZ F1 of 0.89, but only when speech is combined with transcripts. It is the strongest competitor because it is the best published number and it is the representation our method restricts. We run it on identical participant-only audio with the same splits, classifier family and hyperparameter search budget. We will report our own speech-only reproduction rather than rely on 0.89.
Generic acoustic features. The 32-dimensional acoustic features with linear SVM from the Androids Corpus paper. For this baseline we found no reported numeric score on Lacuna; the paper reports only significance over chance. We rerun it with its person-independent five-fold protocol.
Direct physical inversion. The four-parameter two-mass vector fitted per segment by Nelder-Mead, as in the stress-classification paper. This tests whether passing through the encoder helps at all.
Ablations that isolate the key ingredient. The physics-defined directions are compared against four other ways of choosing $k$ dimensions from the same block-8 embeddings at the same $k$:
- the top $k$ principal components of the real data;
- PPA-PCA-PPA reduction, as in Effective Dimensionality Reduction for Word Embeddings;
- a random orthonormal $k$-dimensional subspace;
- a subspace found by PCA on synthetic vowels with tract shape varied and source parameters held fixed (a "tract subspace").
If random or real-data PCA subspaces shrink the transfer gap as much as ours, the benefit is plain dimensionality reduction, not physics.
What counts as beating the strongest method. The physics subspace need not beat full WavLM on DAIC-WOZ in-domain. It wins if its absolute transfer gap to Androids is smaller with non-overlapping speaker-level bootstrap intervals, and if its Androids score after transfer is at least equal.
Why not just…?
Why not just use glottal inverse-filtering measures or standard acoustic features, which are cheaper and already physiological? This is a serious threat. Voice-source measures from inverse filtering may capture most of what the two-mass simulation encodes. The evidence against relying on surface features is suggestive rather than decisive. The neurocomputational source study reports that its source-model features beat OpenSMILE and VATBX on AVEC 2013, and the stress-classification study reports that physical parameters beat F0 and spectral flatness. Neither reports a label-transfer comparison. We therefore include these feature sets in the transfer experiment, at matched dimension, and will report honestly if they close the gap equally well.
Why not just train on the clinician-labelled Androids data directly? We do, as an upper reference. With 118 speakers, though, it cannot support large models. More importantly, almost all available training data are self-report labelled. The practical question is which representations learned from that data survive contact with clinical diagnoses. Training only on Androids cannot answer that.
Why not just invert the physical model per recording and skip the encoder? The inversion papers themselves describe ill-posedness, local minima and sensitivity to noise. Our direct-inversion baseline measures how much this costs on laptop-microphone clinical audio. If direct inversion transfers as well as the subspace, the encoder step is unnecessary and we will say so.
Why not just remove the top principal components as PPA does? That rule is label-free but physics-blind: it removes whatever dominates variance, which may include the very glottal variation that matters. The ablation against PPA-PCA-PPA at equal dimension answers this directly.
Proposed Experiments and Decisive Checks
Grounding check. On FEMH, a classifier on $z$ is compared with the adjoint-inversion accuracies reported in Speech-Based Parameter Estimation of an Asymmetric Vocal Fold Oscillation Model and its Application in Discriminating Vocal Fold Pathologies: 90% normal, 82% neoplasm, 95% phonotrauma and 89% palsy. We do not expect to match these, since they come from a dedicated physical inversion. A physics subspace far below them would signal that synthetic-to-real transfer has failed, and the depression experiments would then be uninterpretable.
Label-source transfer. We train on DAIC-WOZ participant audio (PHQ-8 at or above 10) and separately on AVEC 2013 (BDI-II). We test zero-shot on both Androids tasks, the reading task and the interview task, and also report within-Androids five-fold results. The language difference between English training data and Italian test data affects all representations. Including the tract-subspace control and within-Androids references lets us separate language shift from label-source shift. The reading task, being content-controlled, is the cleanest test.
Prompt-shortcut audit. Using the DAIC-WOZ transcript timing, participant answers are split into responses to distress-triggered follow-ups and responses to neutral prompts, following the shortcut analysis in DAIC-WOZ: On the Validity of Using the Therapist's prompts in Automatic Depression Detection from Clinical Interviews. That paper reports F1 0.88 from interviewer prompts alone and 0.64 from participant text. We measure each representation's accuracy difference between the two answer types. The prediction is near zero for the physics subspace and positive for full WavLM.
Reporting-residual analysis. Within AVEC 2013, we test whether the residual between questionnaire score and physics-subspace prediction correlates with task conditions. The analysis is exploratory and will be labelled as such.
What would refute the proposal. Any of three outcomes:
- a transfer gap for the physics subspace no smaller than that of random or real-data PCA subspaces;
- no reduction in the prompt-type accuracy difference;
- grounding failure on FEMH.
References
- Classification of stressed speech using physical parameters derived from two-mass model
- Neurophysiological Vocal Source Modeling for Biomarkers of Disease
- Deriving Vocal Fold Oscillation Information from Recorded Voice Signals Using Models of Phonation
- Speech-Based Parameter Estimation of an Asymmetric Vocal Fold Oscillation Model and its Application in Discriminating Vocal Fold Pathologies
- Effective Dimensionality Reduction for Word Embeddings
- Functional Principal Component Analysis of Vocal Tract Area Functions
- DAIC-WOZ: On the Validity of Using the Therapist’s prompts in Automatic Depression Detection from Clinical Interviews
- The Androids Corpus: A New Publicly Available Benchmark for Speech Based Depression Detection
- Voice production model employing an interactive boundary-layer analysis of glottal flow
- Self-supervised representations in speech-based depression detection