| The research presented in this dissertation was the result of a concerted effort to evaluate objectively measured features of speech as indicators of vocal affective expression in clinical depression. The purpose of this work was to combine speech signal processing techniques with clinical evaluation of the mental state. Clinical depression was selected for study because it had been identified as the most common mental illness. A collaborative effort with the Psychiatry Department of Behavioral Health (PDBH) at the Medical College of Georgia (MCG) in Augusta was established to help in the collection of a database of patients suffering from depression disorders and a group of controls with no clinical disorders. The patients in the study were from those in the outpatient clinic at MCG. The collaborative effort resulted in the creation of a new speech corpus consisting of a male and female experimental/patient group as well as a male and female control/volunteer group. Speech features related to prosodics (pitch, energy, speaking rate), vocal tract (formant locations, formant bandwidths), and glottal descriptors (glottal timing, glottal ratios, glottal spectrum) were all extracted to create the maximum statistical separation of the control and patient groups, with respect to gender. A robust form of statistical analysis was adopted to identify potential feature statistics for separation. Glottal descriptors proved to be the best separators on a whole for both genders and the use of a set of optimal classifiers produced extremely high separation accuracy at observation lengths as small as 10–20 seconds. The contributions of this study yielded a new speech corpus, a robust feature statistic identification, a new glottal extraction algorithm, and an increased understanding of how effective objectively measured speech features could be at distinguishing vocal affective disorders. |
