In many Western countries depression is a major public health challenge, with its high prevalence and a major impact on patients and economic resources. Effective treatment is available, yet depression care is facing barriers on several levels, such as under-recognition, stigmatization, inadequate treatment and mistreatment. In addition to patient's speech content, doctors can also diagnose depression from the patient's speech quality. They often describe depressed speech as dull, slow and monotonous. We can link these properties with acoustic characteristics.
The aim of my thesis is the separation of healthy and depressed speech with the help of different classification methods. For the classification acoustic, phonetic characteristic parameters are used that reflect changes between healthy and depressed speech. For the examination of the parameters, a processed database containing speech reflecting a depressed mental state, is necessary.
For my paper, I made recordings of patients suffering from depression, at the Department of Psychiatry of Semmelweis Clinic, Péterfy Clinic and Szent János Hospital. Furthermore I made recordings with healthy people for the reference database, keeping in mind to ensure proper age distribution. The participants read a phonetically balanced tale aloud, called “The North Wind and the Sun”. All participants were ranked by a neurologist using a standardized severity scale called BDI.
I labelled and segmented the recordings at phoneme level, then I gained acoustic parameters using Praat. This may be relevant for the classification algorithms for separating healthy and depressed speech. I created a two layered neural network using Matlab and performed the classification. I compared my classification result with SVM classification, which I created with LIBSVM integrated software. The results are encouraging, and legitimize the continuation of the research.