In this paper, we introduce a new set of 3D gesture descriptors based on the labanmovement analysis model. The proposed descriptors are used in a machine learning framework (with SVM and different random forest techniques) for both gesture recognition and emotional analysis purposes. In a first experiment, we test our expressivity model for action recognition purposes on the Microsoft Research Cambridge-12 dataset and obtain very high recognition rates (more than 97 %). In a second experiment, we test our descriptors' ability to qualify the emotional content, upon a database of pre-segmented orchestra conductors' gestures recorded in rehearsals. The results obtained show the relevance of our model which outperforms results reported in similar works on emotion recognition