Speech Features for Depression Detection

Saurabh Sahu, Carol Espy-Wilson

Abstract:

In this paper we discuss speech features that are useful in the
detection of depression. Neuro-physiological changes
associated with depression affect motor coordination and can
disrupt articulatory precision in speech. We use the Mundt
database and focus on six speakers in the database that
transitioned between being depressed and not depressed based
on their Hamilton depression scores. We quantify the degree
of breathiness, jitter and shimmer computed from an AMDF
based parameter. Measures from sustained vowels spoken in
isolation show that all of these attributes can increase when a
person is depressed. In this study, we focused on using
features from free-flowing speech to classify the depressed
state of an individual. To do so we looked at vowel regions
that look the most like sustained vowels. We train an SVM for
each speaker and do a speaker dependent classification of the
test speech frames. Using the AMDF based feature we got a
better accuracy (62-87% frame-wise accuracy for 5 out of 6
speakers) for most speakers than 13 dimensional MFCC along
with its velocity and acceleration coefficients. Using the
AMDF based feature, we also trained a speaker independent
SVM which gave an average accuracy of 77.8% for utterance
based classification.