Controlling a multi-DOF prosthetic hand by EMG signals demands for effective pattern recognition methods that can be easily embedded in the controller of the hand.In this paper,methods of K-nearest neighbor and support vector machine(SVM) were used to identify different modes of myoelectric signals,which were obtained in several on-line experiments.Both methods were performed on different training sample sets,called threshold set and steady-state set,and in the case of abundance and relative insufficiency of samples.Experimental results show that the SVM method is superior to K-nearest neighbor,and the real-time recognition results are better when using threshold dataset as training samples than using steady-state dataset.The proposed method,which is based on SVM and embedded in DSP,can discriminate 10 hand gesture EMG modes with a prediction accuracy of above 95% and a decision frequency of about 30 Hz.