Training four HuBERT-base models on specific aetiologies (CP, PD, and ALS) significantly outperforms mixed-aetiology baselines. Using phonological pseudo-labelling, the models achieved relative F1 improvements of 22.6% for CP, 40.0% for PD, and 32.3% for ALS.
HOW THIS AFFECTS YOU
●
researcherYou can gain significant performance boosts by moving away from heterogeneous label pooling in clinical speech tasks.
●
healthThis improves the accuracy of speech-based diagnostic tools for patients with neurodegenerative diseases.