Analyzing Wav2Vec 1.0 Embeddings for Cross-Database Parkinson's Disease Detection and Speech Features Extraction

Area of Science:

• Speech processing
• Machine learning
• Computational linguistics

Background:

• Deep learning advancements enable Parkinson's disease (PD) modeling using extensive unlabeled speech data.
• Minimal annotated data is required for effective PD speech analysis.
• Non-fine-tuned wav2vec 1.0 architecture offers a novel approach for PD speech modeling.

Purpose of the Study:

• To develop machine learning models for PD speech diagnosis using wav2vec 1.0.
• To analyze overlapping components within embeddings for classification and regression tasks.
• To investigate shared latent speech representations in PD across different models and tasks.

Main Methods:

• Utilized the non-fine-tuned wav2vec 1.0 architecture for PD speech modeling.
• Evaluated models on three multi-language PD datasets for cross-database classification.
• Employed regression tasks to predict demographic and articulation characteristics.
• Analyzed feature importance to identify shared components between classification and regression models.

Main Results:

• Wav2vec accurately detected PD, outperforming mel-frequency cepstral coefficients in cross-database classification.
• Cross-database performance with wav2vec was comparable to intra-dataset evaluations.
• Wav2vec effectively modeled speech characteristics related to articulation and aging.
• Significant feature overlap was found between classification and regression models, indicating improved generalizability.

Conclusions:

• Wav2vec embeddings show promise for accurate and generalizable PD speech diagnosis.
• Shared features across related tasks suggest the potential for a universal speech-based PD evaluation model.
• This approach facilitates PD assessment using extensive unlabeled speech data with minimal annotation.