Comparison of read and spontaneous speech in case of automatic detection of depression

摘要

In this paper, read and spontaneous speech have been compared in the light of automatic depression detection by speech processing. First, statistical analysis was carried out to select those acoustic features that differ significantly between healthy and depressed subjects in case of these two types of speech, separately for both gender. Secondly, statistical examination and classification experiments were prepared to compare the values of the selected features for the two types of speech. We were looking for the answer to which type of speech can be used to achieve better automatic depression detection results. As it was expected, the tempo related features, such as articulation rate, speech rate, and pause lengths are useful in case of spontaneous speech, while formants trajectories can be used only in case of read speech, because their values are mainly influenced by the linguistic content of the speech. Despite the significant differences of the features' values between read and spontaneous speech, there were no major differences in the detection accuracies. 83% detection accuracy was archived with read speech samples, and 86% detection accuracy was achieved with spontaneous speech samples.