On Speech Datasets in Machine Learning for Healthcare

Abstract

Multi-language speech datasets are scarce and often have small sample sizes in the medical domain. We address this problem by employing the cross-linguistic transfer methods and by collecting the large longitudinal dataset of impaired speech. The cross-lingual method demonstrates improvements in Aphasia detection over unilingual baselines, and the early results on the newly collected dataset show the promise to achieve a strong baseline in Alzheimer’s disease detection