| Popis: |
Numerous changes in health condition can cause speech disorders and voice impairments. The present paper aims not only to separate speech samples of healthy speakers from a larger set of diseases, namely depression, Parkinson’s disease, structural organic dysphonia, functional dysphonia and recurrent paresis, but to examine the classification possibilities of these diseases. Datasets containing read texts were used. Support vector machines, neural networks and k-nearest neighbor methods were utilized to compare performances. Flat and hierarchical classification scenarios were also investigated, followed by a final feature selection step. The best overall accuracy was 66.1% using six classes and 88.0% using re-categorization of disorders into four classes. Flat classification scenario outperformed the hierarchical one. |
