Fair Compression of Speech Features for Depression Detection, Anchored to Symptoms Reported the Same Way Across Groups

Speech-based depression detectors are trained on tiny datasets. DAIC-WOZ, the standard benchmark, has fewer than 200 clinical interviews. So practitioners almost always compress thousands of acoustic measurements into a few dozen directions before fitting a classifier, typically with principal component analysis (PCA), which keeps the directions along which the pooled data vary most. Because male and female voices differ so much, those high-variance directions are largely about who is speaking, not whether they are depressed. Whatever depression signal the minority group's voices carry along other directions is thrown away before any fairness-aware classifier ever sees it. A second problem compounds the first. The labels are self-reported questionnaire scores (the PHQ-8), and if men and women answer some items differently for the same underlying severity, a model made "fair" against those labels inherits the reporting bias.

We propose a compression step that, for each demographic group separately, keeps the acoustic directions that predict depression symptoms, and that maximizes the worst group's share. The symptoms it is anchored to are restricted to questionnaire items that pass a test showing they mean the same thing across groups. In two sentences: choose the compressed subspace so that the group it serves worst still retains as much symptom-predictive signal as possible, rather than so that pooled variance is maximal. Define "symptom" only through questionnaire items that both groups demonstrably interpret alike, so the fairness target is not itself a reporting artifact.

Core Claim

Existing fair dimensionality reduction equalizes the wrong quantity for this problem. Multi-Criteria Dimensionality Reduction with Applications to Fairness shows that standard PCA imposes a "tyranny of the majority," with minority reconstruction error nearly double that of a fair alternative on UCI credit data. Its Fair-PCA fixes this by minimizing the worst group's reconstruction error, solved as a semidefinite program (an optimization over positive semidefinite matrices) that is exact for two groups. Reconstruction error, however, rewards keeping pitch and formant variance that is irrelevant to depression. We claim that replacing per-group reconstruction with per-group label-relevant variance, computed against measurement-invariant symptom items, reduces worst-group detection error relative to both standard PCA and Fair-PCA at equal dimensionality. We further claim that part of the gender gap measured on raw PHQ-8 totals is a labeling artifact that disappears when evaluation is restricted to invariant items. Either claim can fail, and the experiments are designed so that each failure is visible.

The mechanism needs one equation. Pool and whiten the acoustic features (with ridge shrinkage), and let $\Sigma_{xy,g}$ be the cross-covariance between whitened features and the retained PHQ-8 items in group $g$. Set $B_g = \Sigma_{xy,g}\Sigma_{yx,g}$. In whitened coordinates, $\mathrm{tr}(B_g P)$ is the item variance that a linear map from a rank-$d$ subspace with projector $P$ can explain in group $g$. We solve

$$ \min_{P}\ \max_{g}\ \Big(\beta_g^{*}(d) - \mathrm{tr}(B_g P)\Big) \quad \text{s.t. } \mathrm{tr}(P)\le d,\ 0\preceq P\preceq I, $$

where $\beta_g^{*}(d)$ is the best value group $g$ could attain with its own rank-$d$ subspace, so each group's loss is measured against its own ceiling. The objective is linear in $P$ with the same constraint set as Fair-PCA. The method of Multi-Criteria Dimensionality Reduction with Applications to Fairness therefore applies directly: an exact rank-$d$ solution for two groups and iterative rounding for more. Its released SDPforAll/multiCriteriaDimReduction code is reused with $B_g$ substituted for the group covariance matrices. That paper supplies the solver; what is new here is the objective and the label definition.

Relation to Prior Work

The closest fairness work in depression detection is FairReFuse. It removes gender confounding through causal backdoor adjustment and reweights audio against video, reaching F1 0.70 against a 0.59 baseline and an equal-opportunity ratio of 1.02 on D-Vlog. It operates on dense, uncompressed embeddings and treats PHQ-8 as ground truth. We differ on both counts: we act at the compression stage, and we audit the labels. Our method is not a replacement for FairReFuse. Its debiased encoder could consume our subspace, and we test it as a competitor.

Three papers motivate compression in this regime and constrain how it is evaluated. Interspeech Pathology Challenge: Investigations into Speaker and Sentence Specific Effects shows accuracy collapsing from 85.5% to 59.0% when speakers are kept out of training. Projecting about 6,000 descriptors to 50 principal components partially recovers it, to 66.9% development and 64.14% evaluation weighted accuracy. That result motivates compression and requires speaker-exclusive splits. Self-Supervised Representations in Speech-Based Depression Detection finds the middle layers (block 8) of speech foundation models most useful and documents overfitting on DAIC-WOZ. That tells us which embedding to compress. Crosslinguistic Acoustic Feature-based Dementia Classification Using Advanced Learning Architectures applies 95%-variance PCA to eGeMAPSv2 and ComParE across English, Korean, and Mandarin and reports 83.54% accuracy and AUC 0.80. It never asks whether the pooled projection serves each language equally, which makes it a natural second testbed with three groups.

The label side builds on two papers. Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders demonstrates multi-group confirmatory factor analysis across configural, metric, scalar, and strict tiers ($\Delta$CFI $\le 0.002$ with 1,120 respondents). Confirmatory factor analysis tests whether questionnaire items load on the same latent trait with the same weights and thresholds in each group. We borrow this procedure for PHQ-8 items. Predicting Individual Depression Symptoms from Acoustic Features During Speech shows that speech maps unevenly to individual items and that predicting items jointly helps. That justifies using the item vector rather than the total as the compression target. Unlike that work, we compress, and we stratify by group.

One constraint shapes the inputs. DAIC-WOZ: On the Validity of Using the Therapist's prompts shows that interviewer prompts alone reach F1 0.88 in text. We therefore use only participant speech segments, so the compressed subspace cannot encode the interviewer's behavior.

Research Program

The first stage is a label audit. On DAIC-WOZ and E-DAIC, we fit a one-factor model to the eight PHQ-8 items and test invariance by gender, tier by tier. Where invariance fails, we free individual item thresholds or loadings until partial invariance holds. The sample is small, so we report bootstrap intervals on the fit changes and treat an item as non-invariant only when that conclusion survives resampling. The output is a set of invariant items and an invariant subscore. If every item passes, that is itself a finding: on this benchmark, gender disparities cannot be attributed to the questionnaire, and the second claim fails cleanly.

The second stage is the compression comparison. Two feature sets are used: the 88-dimensional eGeMAPS set, and mean-pooled WavLM block-8 embeddings. The WavLM embeddings are pre-reduced only by pooled ridge whitening. We compute $B_g$ per group on training folds only and solve the program above for $d$ between 5 and 50. We fit the same downstream model on every compressed representation. That model is a ridge multi-output regressor over items, plus a logistic head for the PHQ-8 cutoff of 10.

The third stage is a transfer check on the three-language dementia corpora. There the groups are languages ($k=3$, via iterative rounding), the label is AD versus healthy control, and $B_g$ reduces to the between-class scatter within each language.

Baselines

The naive reference point is no compression: the same ridge and logistic models on full eGeMAPS or whitened WavLM features, under the same speaker-exclusive splits. If this matches our method on worst-group error, compression is not where bias enters and the premise fails.

The conventional pipeline is pooled PCA at matched $d$, set by the 50-component recipe of Interspeech Pathology Challenge and by the 95%-variance recipe of Crosslinguistic Acoustic Feature-based Dementia Classification. On the dementia data, beating it means at least matching the reported 83.54% accuracy and AUC 0.80 while raising the worst language's AUC. Any deviation of our re-run of their pipeline from those figures will be reported.

Fair-PCA from Multi-Criteria Dimensionality Reduction with Applications to Fairness, run with the same solver, isolates label-awareness. The only difference from our method is reconstruction covariance versus $B_g$. A pooled label-aware variant, which maximizes $\mathrm{tr}(\sum_g B_g P)$ (ordinary reduced-rank regression), isolates the minimax-over-groups ingredient. Running our method with all eight items versus invariant items only isolates the label audit.

The strongest existing fairness method is FairReFuse, restricted to its audio branch for comparability. It reports F1 0.70 and equal opportunity 1.02 on D-Vlog. The file contains no DAIC-WOZ or E-DAIC breakdown, so no reported result was found on Lacuna for FairReFuse on DAIC-WOZ, and we re-run it ourselves on identical folds. We also stack it on our subspace to test complementarity. For context on accuracy, Self-Supervised Representations in Speech-Based Depression Detection reports F1 0.89 for a speech-plus-text ensemble on DAIC-WOZ. We do not expect a compressed linear audio-only model to reach that figure, and we do not claim to.

Fairness requires an equal budget. Every method gets the same folds, the same downstream model, and the same grid over $d$ and ridge strength, selected on inner folds by worst-group validation error. Beating a baseline means a lower worst-group PHQ-8 mean absolute error and a smaller gender gap in true-positive rate, without a statistically detectable loss in pooled F1, with intervals computed across repeated speaker-exclusive splits.

Why not just…?

Why not just standardize features within each gender before ordinary PCA? This is the most serious threat. Much of the voice difference between sexes is a shift in mean and scale, and per-group z-scoring removes it cheaply. Our method helps only if groups differ in which directions carry symptom information, not merely in offsets. We include per-group standardization plus PCA as a baseline, and we also apply our method on top of standardized features. If standardization alone closes the worst-group gap, we will report that the simple fix suffices and that the remaining contribution is the label audit.

Why not just use Fair-PCA as published? Equal reconstruction error can be achieved by preserving each group's pitch and formant variance, which has nothing to do with depression. The covariate analysis in Acoustic characteristics of depression in older adults' speech shows that depression markers such as F2 range and pause duration are small effects layered under demographic variance. The Fair-PCA ablation tests this directly. If Fair-PCA performs as well, label-awareness adds nothing.

Why not skip compression and fix fairness at the output with threshold adjustment or an adversarial head? Output-side methods cannot recover directions that were never seen. Without compression, however, nothing is discarded, and the remaining question is overfitting. The no-compression baseline with post-hoc equalized-odds thresholds tests whether that combination matches our method. FairReFuse covers the strongest learned version of this alternative. Stark et al.'s speaker-exclusive collapse suggests overfitting will hurt the uncompressed route on DAIC-WOZ-sized data, but we measure rather than assume.

Why not train a separate model per group? Each gender in DAIC-WOZ has only a few dozen depressed participants, so per-group models should have high variance. We include per-group PCA with a per-group classifier as a baseline. It is also the natural upper reference for $\beta_g^{*}(d)$.

Decisive Checks

Three outcomes decide the proposal. First, at matched $d$ on participant-only DAIC-WOZ audio, our method must beat both per-group-standardized PCA and Fair-PCA on worst-group PHQ-8 error, with the gap persisting across repeated speaker-exclusive splits. Second, the label audit must identify at least one non-invariant PHQ-8 item. Restricting training targets to invariant items must then shrink the gender gap measured on the invariant subscore more than it shrinks pooled accuracy. If the gap shrinks only when measured on the full total, the effect is a labeling artifact and should be reported as such. Third, on the three-language dementia data, the method must raise the worst language's AUC above pooled PCA's without lowering pooled accuracy below the reported 83.54%. That would show the mechanism is not specific to gender or to self-report labels.

If the first check fails while the second succeeds, the contribution narrows to a measurement finding about DAIC-WOZ labels. If the first succeeds and the second fails, it is a compression method for clinical speech with a clean negative result on the questionnaire.