Our work allows an interactive music system to spot and recognize "command" gestures from musicians in real time. The system gives the musician gestural control over sound and the flexibility to make distinct changes during the performance by interpreting gestures as discrete commands. We combine a gesture threshold model with a Dynamic Time Warping (DTW) algorithm for gesture spotting and classification. The following problems are addressed: i) how to recognize discrete commands embedded within continuous gestures, and ii) an automatic threshold and feature selection method based on F-measure to find good system parameters according to training data.