Parkinson’s disease (PD) is a neurodegenerative disease that produces progressive motor impairments. Dysarthria (speech disorders) and hypomimia (face rigidity) are two major Parkinsonism patterns observed even at the early stages of the disease. Nonetheless, the clinical diagnosis is mainly observational and dependent on the specialists’ expertise. Besides, the categorization of each of these patterns is isolated, which may lead to delayed diagnosis and misplanning of treatments. This work introduces a non-invasive multimodal strategy that integrates video and audio modalities into the online characterization of speech exercises. Subjects were invited to pronounce sustained vowels while video and audio were recorded. Then, a temporal window is run along the sequence to build online covariance matrices of synchronized face landmarks position and characteristic voice frequencies. From these temporal covariance matrices are learned Riemannian descriptors that allow to discriminate between Parkinson’s and control subjects. From a study with 14 subjects, the proposed approach achieved a mean accuracy of 70% in sustained vowel pronunciation. Considering online predictions, the proposed approach evidenced a consistent accuracy of 0.77 during pronunciation of close vowels.