In this paper we presented a speed analysis of arm movement. Results show that 1) instantaneous velocity provides more reliable classification compared to instantaneous acceleration, 2) the mean is a better feature compared to the standard deviation for instantaneous velocity, and 3) the hand joint is the most efficient joint for speed detection in arm motion. Moreover, a low-pass filter improves the interclass speed classification but has no effect on the intraclass classification. For interclass speed classification, the mean of non-filtered instantaneous velocities scored a 0.49% error rate in detecting the speed type over 405 motions, while the standard deviation of filtered instantaneous velocity scored 5.43% during the intraclass classification. Moreover, the first 60% of the range of movement provides a classification error relatively close to the use of the whole movement range—this can be used for predicting the speed type in real time. Furthermore, the most important 10% of a whole motion is the 50–60% region. The results are promising, and this approach can be implemented in a human–computer intraction system for interactive tremor diagnosis, specifically measuring hand-related disability and improvement. In this study, we asked healthy subjects to mimic abnormality by moving slowly; however, testing this approach on patients with Parkinson’s disease or any hand tremors remains a task for future work.