Language Orthogonalization of Self-Supervised Speech Representations for Cross-lingual Parkinson's Detection

I. INTRODUCTION

• Speech-based PD biomarkers are being investigated because motor-speech impairments can be captured non-invasively, repeatedly, and at low cost.S3Ms encode acoustic and phonetic information and have outperformed handcrafted features in dysarthria and PD detection.Source: §I · p. 1
• When target-language training speech contains only HC speakers, language-dependent S3M structure can lead classifiers to label unseen target-language patients as HC.On OneVoice-MSD-2026, the resulting bias produces high specificity but low sensitivity, a concern for screening because missed patients are false negatives.Source: §I · p. 1
• Unlike language shift, which translates each language by a fixed vector and changes its mean, language orthogonalization estimates speaker-varying language structure.This allows it to reduce language-dependent variation beyond the distribution mean.Source: §I · p. 1
• Language orthogonalization fits an HC-only ridge regression from external language-identification embeddings to S3M representations, then subtracts the predicted component.Fitting on HC speech avoids directly modeling PD-associated variation; the residual is intended to reduce language separation while retaining pathology-related variation.Source: §I · p. 1
• The paper reports consistent improvements across five S3M backbones, three speech tasks, and three target languages, including sensitivity-threshold transfer and reduced language decodability.The schematic depicts HC representations clustering while PD representations spread outward in the residual space.Source: §I · p. 2, Figure 1 · p. 2

II. METHOD

• The pipeline extracts S3M and external language-identification embeddings, then subtracts the S3M component predictable from the language embedding to retain a residual.The residual is expected to reduce language separation while retaining pathology-related variation.Source: §II · p. 2

A. S3M backbones and pooling

• Five frozen 24-layer Transformer-Large encoders produce 2048-dimensional utterance embeddings by averaging normalized mean-and-standard-deviation pooled representations across all 25 depths.Speaker representations average utterance embeddings within each task; the same pipeline is used across backbones rather than tuning a layer choice.Source: A. S3M backbones and pooling · p. 2

B. External LID embeddings

• VoxLingua107 ECAPA-TDNN supplies 256-dimensional speaker-level language embeddings, averaged across utterances within each speaker and task.The 14M-parameter encoder was pretrained to identify 107 languages and does not rely on PD labels.Source: B. External LID embeddings · p. 2

C. Alignment methods

• LS aligns each language’s training-fold HC centroid with the target-language HC centroid by applying a constant offset to that language.Source: C. Alignment methods · p. 2
• LO fits a ridge map from language-identification embeddings to S3M representations and subtracts the language-predictable component.The map is fitted on HC speakers; the resulting residual is used for downstream classification.Source: C. Alignment methods · pp. 2–3
• Smaller α values remove more LID-predictable variation; results are reported across α values.The ridge objective uses the HC S3M and LID representations to fit the map.Source: C. Alignment methods · pp. 2–3
• Unlike LS, which subtracts one fixed vector per language, LO subtracts a speaker-varying language-predictable component and can reduce variation beyond the distribution mean.Both alignment methods are fitted on the HC subset and applied to all training and evaluation speakers.Source: C. Alignment methods · pp. 2–3

III. EXPERIMENTAL SETUPS

• OneVoice-MSD-2026 harmonizes GermanPD (n=176), CzechPD (n=100), and Spanish e-PC-GITA (n=140) under a common protocol.Source: A. Dataset · p. 3
• Each speaker performs sustained vowel phonation, oral diadochokinesis, and read speech, with binary HC-versus-PD labels.DDK is rapid /pa-ta-ka/ repetition.Source: A. Dataset · p. 3
• Each language is split into five speaker-disjoint folds, stratified jointly by group and language to preserve HC/PD balance within every language.Source: A. Dataset · p. 3

B. Cross-lingual protocol

• Training uses all source-language speakers and target-language HC folds except the held-out fold; evaluation uses that HC fold and all target-language PD speakers.Source: B. Cross-lingual protocol · p. 3
• PD detection uses class-balanced logistic regression, with features standardized using training-set statistics only.Source: B. Cross-lingual protocol · p. 3

C. Metrics

• The primary metric is F1 at the 0.90-sensitivity operating point; sensitivity achieved on target-language patients is also compared with the sensitivity requested during training.The F1 metric is averaged over five backbones, three tasks, three target languages, and five folds; Table I averages over five backbones and three target languages.Source: C. Metrics · p. 3, Table I · p. 3
• Figure 2 plots cross-lingual F1 against α for five backbones and three tasks, averaged over three target languages.The shaded region marks raw and +LS baselines, and stars mark each curve's maximum F1.Source: Figure 2 · p. 3

• The results test whether the method improves cross-lingual PD detection over the baselines.Source: §IV · p. 3
• LO closely follows requested sensitivities from 0.70 to 0.90 on the held-out target language, whereas raw and LS fall well below the diagonal.Source: Figure 3 · p. 4
• The section focuses on high-specificity, low-sensitivity performance, threshold transfer, and application to screening.Source: §IV · p. 4

A. Cross-lingual PD detection

• At the 0.90-sensitivity operating point, averaged F1 was 0.87 for LO (α = 0), versus 0.52 for raw features and 0.67 for LS, a +0.35 gain over raw.Figure 2 shows LO outperforming both baselines over a broad range of α.Source: A. Cross-lingual PD detection · p. 4

B. Threshold transferability

• With LO, achieved target-language sensitivity closely follows the requested value across 0.70–0.90, unlike raw and LS; thresholds are selected using training-set predictions.This enables threshold selection without target-language PD data for calibration.Source: B. Threshold transferability · p. 4

C. Application to PD screening

• LO reaches the screening region across target languages, while raw and LS remain in the high-specificity, low-sensitivity region despite screening-selected thresholds.LO's projection in the read-speech German example shows HC representations concentrated near the center and PD representations more dispersed.Source: Figure 4 · p. 4, C. Application to PD screening · p. 5, Figure 5 · p. 5

A. Representation geometry

• HC speakers form a tight, low-norm baseline, whereas PD speakers are distributed farther from the residual centroid.The paper relates this pattern to the heterogeneous nature of dysarthric deviations in PD speech.Source: A. Representation geometry · p. 5
• Language orthogonalization produces greater overlap among language groups than language shift, which aligns language means but preserves within-language covariance.LS translates every representation from a language by the same vector; LO subtracts a speaker-specific prediction.Source: A. Representation geometry · p. 5
• Across tasks, PD speakers’ median residual-centroid distance is 7.5–10.8 times that of HC speakers.Figure 6 measures each speaker’s orthogonalized representation against the residual centroid.Source: A. Representation geometry · p. 5, Figure 6 · p. 5

B. Residual language information

• LO yields language-classification macro-F1 scores of 0.47 for HC and 0.62 for PD, the lowest scores for both groups among the transformations.Some language information remains, particularly among PD speakers.Source: B. Residual language information · p. 6
• Raw representations yield language-classification macro-F1 scores of 0.99 for HC and 0.98 for PD.The classifier predicts three languages; chance performance is 0.33.Source: B. Residual language information · pp. 5–6
• LS reduces language-classification macro-F1 to 0.76 for HC and 0.94 for PD, reducing decodability less than LO.Source: B. Residual language information · p. 6

C. Applicability beyond S3Ms

• In exploratory German read-speech projections, LO increased language overlap in ECAPA-TDNN, Whisper, and AST representations while HC and PD speakers remained differently distributed.These findings motivate more comprehensive evaluation beyond S3Ms.Source: C. Applicability beyond S3Ms · p. 6, Figure 7 · p. 6

VI. DISCUSSION

• Evaluation covered Germanic, Slavic, and Romance languages within Indo-European; testing more typologically diverse languages remains an open need.The authors note that LO may behave differently as representational distance between source and target languages increases.Source: §VI · p. 6
• The authors identify task dependence, nonlinear extensions, cross-language calibration, and target-free α selection as directions for further work.LO showed larger gains for sustained vowels than read speech; future work could also examine specificity while maintaining screening sensitivity.Source: §VI · p. 6

VII. CONCLUSION

• Across five S3M backbones, three tasks, and three target languages, LO improved F1 over raw features and LS and mitigated the high-specificity, low-sensitivity failure on target-language patients.The conclusion also reports improved transfer of sensitivity-targeted thresholds without target-language PD data.Source: §VII · p. 6