Exploring transfer learning for pathological speech feature prediction: Impact of layer selection

One approach to There is interest in leveraging AI to conduct automatic, objective address this is to focus on the pathological speech features that assessments of clinical speech, in turn facilitating diagnosis characterize speech disorders and create a model that predicts and treatment of speech disorders. We explore transfer them instead of predicting disease [9]. The key insight behind learning, focusing on the impact of layer selection, for this approach is that there are predictable mappings between the downstream task of predicting the presence of pathological groupings of features and disorder, and not all features speech. We find that selecting an optimal layer offers large are unique with respect to disease [10, 8]. Because of this, performance improvements ( 12.4% average increase in balanced a model could learn the information necessary to recognize a accuracy), though the best layer varies by predicted feature specific type of dysarthria using recordings from a cohort with and does not always generalize well to unseen data. A varying neurological diseases which can cause that dysarthria.