Speaker-independent dysarthria severity classification using self-supervised transformers and multi-task learning.

Main Methods:

• Utilized a transformer-based architecture integrating a wav2vec 2.0 model pre-trained on healthy speech data.
• Implemented a contrastive learning strategy with a multi-task objective: cross-entropy loss for severity classification and triplet margin loss for speaker-agnostic embedding.
• Employed Speaker-Agnostic Latent Regularisation (SALR) to group latent speech embeddings by severity, not by speaker.

Main Results:

• The SALR framework achieved 70.5% accuracy and 59.2% F1 score on the UA-Speech dataset using leave-one-subject-out cross-validation.
• Demonstrated a significant improvement of 16.5% absolute (30% relative) over previous benchmarks.
• Explainability analysis confirmed enhanced ordinal structure in the latent space, reducing speaker-specific dependencies and showing robustness.

Conclusions:

• The SALR framework shows significant potential for speaker-independent dysarthria severity classification.
• This approach offers an objective, accessible, and cost-effective alternative to traditional dysarthria assessment methods.
• The SALR framework has promising implications for automated clinical applications in speech disorder assessment.