In this paper, we propose a method to discriminate fingers activity and wrist activity from electromyogram (EMG) picked up from surface electrode. Generally it is difficult to pick up the EMG that comes from muscle fibers of the finger flexors because they are located at the substantially deep layer of the forearm. We focus on difference between the EMG that comes from the muscle fibers located in the deep layer and the shallow layer in two fold. One is the difference of frequency distribution, and the other is the difference of their EMG amplitude due to whether it is picked up by the wide distance differential electrodes or by the narrow distance ones. We introduce the Recursive Running DFT to discriminate on the first feature on real-time, and we also develop a double surface differential electrode to pick up two channels EMG with different length of electrodes simultaneously to find the second feature. We found the combination of these features allows us to discriminate a fingers activity, a wrist activity and their combined activity. The experiments with five subjects revealed the effectiveness of the proposed method. The method is calculated so fast that it is expected to serve it for real-time control of multi-functional forearm prosthesis.