Background: Parkinson's disease patients (PDP) are evaluated using the unified Parkinson's disease rating scale (UPDRS) to follow the longitudinal course of the disease. UPDRS evaluation is performed by a neurologist, and hence its use is limited in the evaluation of short-term (daily) fluctuations. Subjects taking L-DOPA as part of treatment to reduce symptoms exhibit motor fluctuations as a common complication. Objectives: The aim of the study is to assess the use of speech analysis as a proxy to continuously monitor PDP medication state. Methods: We combine acoustic, prosody, and semantic features to characterize three speech tasks (picture description, reverse counting and diadochokinetic rate) of 25 PDP evaluated under different medication states: ON and OFF L-DOPA. Results: Classification of medication states using features extracted from audio recordings results in cross-validated accuracy rates of 0.88, 0.84 and 0.71 for the picture description, reverse counting and diadochokinetic rate tasks, respectively. When adding feature selection and semantic features, the accuracy rates increase to 1.00, 0.96 and 0.83 respectively; thus reaching very high classification accuracy on 3 different tasks. Conclusions: We show that speech-based features are highly predictive