Analysis of Prosodic Features and Formant of Dementia Speech for Machine Learning

In previous research of dementia discrimination by voice, the discrimination method using multiple acoustic features by machine learning has been proposed. However, they do not focus on speech analysis in patients with mild dementia. We analyzed the prosodic features and Formant of vowels that have not yet been analyzed. As a result of the t-test, it was confirmed that the value of Jitter and HNR tended to be high, and the parameters of F0 and Shimmer tended to be low. In addition, machine learning discrimination revealed an F-score of 0.621 (SVM), suggesting that prosodic features and formant alone are insufficient.