Emotional Temperature for the Evaluation of Speech in Patients with Alzheimer's Disease through an Automatic Interviewer

Abstract

In the context of the detection and evolutionary control of Alzheimer's disease from voice recordings and their automatic processing, this work aims to objectively determine the discriminatory capacity of a set of voice features linked to the emotional load of speech. We use descriptive statistics derived from the concept of emotional temperature as quantifiable characteristics of the voice. We apply a series of parametric and nonparametric analyses to the set of features, both individually and collectively, and explore their potential in relation to the use of different methods of unsupervised classification. With the aim of comparing how the type of interviewer used in the sample collection (i.e., voice recordings) influences the discrimination of AD through emotional speech analysis, we used the CSAP-19 database, which includes voice samples obtained through human interviewer (spontaneous speech samples) and automatic interviewer (induced speech samples) for the three defined populations (HC, mild AD, and moderate AD). In this regard, a comparative analysis is also conducted on the potential of emotional temperature features defined according to the sample collection process (manual or automatic interview process).