Self-Supervised Speech Representations for Cross-Speaker Dysarthria Detection During Awake Craniotomy
The study evaluated a pipeline for detecting dysarthric speech during awake craniotomies using patient recordings. The pipeline combines speaker diarization, multi-view representations (acoustic features and wav2vec 2.0 embeddings), and a cascaded classifier. Results show that the quality of the speech representation, especially using self-supervised learning embeddings, is far more important for cross-speaker performance than the choice of classifier.