Abstract
In this paper class posterior distributions are combined with a hierarchal structure of multilayer Perceptrons to perform an automatic assessment of dysarthric speech. In addition to the standard Mel-frequency coefficients, this hybrid classifier uses rhythm-based features as input parameters since the preliminary evidence from perceptual experiments show that rhythm troubles may be the common characteristic of various types of dysarthria. The Nemours database of American dysarthric speakers is used throughout experiments. Results show the relevance of rhythm metrics and the effectiveness of the proposed hybrid classifier to discriminate the levels of dysarthria severity.