This paper deals with Parkinson's disease (PD) severity estimation according to the Unified Parkinson's Disease Rating Scale: motor subscale (UPDRS III), which quantifies the hallmark symptoms of PD, using an acoustic analysis of speech signals. Experimental dataset comprised 42 speech tasks acquired from 50 PD patients (UPDRS III ranged from 6 to 92). It was divided into subsets: words, sentences, reading text, monologue and diadochokinetic tasks. We performed a parametrization of the whole corpus and these groups separately using a wide range of conventional and novel speech features. We used guided regu-larized random forest algorithm to select features with maximum clinical information and performed random forests regression to estimate PD severity. According to significant correlations between true UPDRS III scores and scores predicted by the proposed methodology it was shown that information extracted through variety of speech tasks can be used to estimate the degree of PD severity.