In this paper we present a method for the recognition of dynamic gestures with discrete Hidden Markov Models (HMMs) from a continuous stream of gesture input data. The segmentation problem is addressed by extracting two velocity profiles from the gesture data and using their extrema as segmentation cues. Gestures are captured with a TUB-SensorGlove. The paper focuses on the description of the gesture recognition method (including data preprocessing) and describes experiments for the evaluation of the performance of the recognition method. The paper combines and further develops ideas from some of our previous work.