Speech biomarkers for automated depression level detection

Abstract

This study investigates the contribution of speech audio and speech verbal content in the automated detection of depression levels. Recordings from the Distress Analysis Interview Corpus Wizard-of-Oz dataset and the depression severity labels of the recordings were used to extract acoustic features. A transcription of the recordings was used to extract textual features. The acoustic set included prosodic, cepstral, and glottal feature categories. The textual features consisted of semantic and syntactic categories. Mutual information feature selection, followed by a random forest classifier identified the set of features which optimised the depression level classification. The optimised binary classification of depression from non-depressed yielded an accuracy of 0.89 and an F1 score of 0.87. A classification of the five depression levels yielded an accuracy of 0.79 and an F1 score of 0.72. The ratio of importance scores of acoustic to textual of the speech acoustic features was greater than 3:1. Our method thus provided acoustic and textual indicators in depressed speech. These might increase the acceptability of automated depression detection by healthcare professionals. Our initial findings indicate a select set of features that can improve the effectiveness of automated depression detection and monitoring tools.