Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bilingual Dual-Head Deep Model for Parkinson's Disease Detection from Speech

Moreno La Quatra, Juan Rafael Orozco-Arroyave, Marco Sabato Siniscalchi

TL;DR

This work tackles Parkinson's disease detection from speech in a bilingual setting by introducing a dual-head deep architecture with task-specific branches for DDK and continuous speech. A shared backbone processes self-supervised learning (SSL) features and wavelet-based representations, augmented by adaptive layers, convolutional bottlenecks, and contrastive learning to reduce cross-language variability. Experimental results on Slovak (EWA-DB) and Spanish (PC-GITA) show that naive multilingual training fails to generalize, while the proposed bilingual model achieves high accuracy in both languages and superior cross-language generalization. The findings highlight a scalable approach for robust PD screening across diverse linguistic contexts, with implications for early diagnosis and widespread deployment.

Abstract

This work aims to tackle the Parkinson's disease (PD) detection problem from the speech signal in a bilingual setting by proposing an ad-hoc dual-head deep neural architecture for type-based binary classification. One head is specialized for diadochokinetic patterns. The other head looks for natural speech patterns present in continuous spoken utterances. Only one of the two heads is operative accordingly to the nature of the input. Speech representations are extracted from self-supervised learning (SSL) models and wavelet transforms. Adaptive layers, convolutional bottlenecks, and contrastive learning are exploited to reduce variations across languages. Our solution is assessed against two distinct datasets, EWA-DB, and PC-GITA, which cover Slovak and Spanish languages, respectively. Results indicate that conventional models trained on a single language dataset struggle with cross-linguistic generalization, and naive combinations of datasets are suboptimal. In contrast, our model improves generalization on both languages, simultaneously.

Bilingual Dual-Head Deep Model for Parkinson's Disease Detection from Speech

TL;DR

This work tackles Parkinson's disease detection from speech in a bilingual setting by introducing a dual-head deep architecture with task-specific branches for DDK and continuous speech. A shared backbone processes self-supervised learning (SSL) features and wavelet-based representations, augmented by adaptive layers, convolutional bottlenecks, and contrastive learning to reduce cross-language variability. Experimental results on Slovak (EWA-DB) and Spanish (PC-GITA) show that naive multilingual training fails to generalize, while the proposed bilingual model achieves high accuracy in both languages and superior cross-language generalization. The findings highlight a scalable approach for robust PD screening across diverse linguistic contexts, with implications for early diagnosis and widespread deployment.

Abstract

This work aims to tackle the Parkinson's disease (PD) detection problem from the speech signal in a bilingual setting by proposing an ad-hoc dual-head deep neural architecture for type-based binary classification. One head is specialized for diadochokinetic patterns. The other head looks for natural speech patterns present in continuous spoken utterances. Only one of the two heads is operative accordingly to the nature of the input. Speech representations are extracted from self-supervised learning (SSL) models and wavelet transforms. Adaptive layers, convolutional bottlenecks, and contrastive learning are exploited to reduce variations across languages. Our solution is assessed against two distinct datasets, EWA-DB, and PC-GITA, which cover Slovak and Spanish languages, respectively. Results indicate that conventional models trained on a single language dataset struggle with cross-linguistic generalization, and naive combinations of datasets are suboptimal. In contrast, our model improves generalization on both languages, simultaneously.

Paper Structure

This paper contains 13 sections, 10 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method
  4. Type-Based Classifiers
  5. Wavelet Integration
  6. Adaptive Layers
  7. Bottleneck Layers
  8. Contrastive Learning
  9. Experiments
  10. Datasets
  11. Experimental Setup and Metrics
  12. Results and Analysis
  13. Conclusions

Figures (3)

  • Figure 1: Proposed model architecture. Each branch includes an attention pooling layer (AP) and two linear classification layers with ReLU. HC refers to Healthy Control, PD to Parkinson's Disease, and $\sigma$ is the sigmoid activation.
  • Figure 2: Ablation study results showing the impact of removing individual components on the F1 score for EWA-DB and PC-GITA datasets.
  • Figure 3: t-SNE visualization of DDK task embeddings.