Brain decoding would be a replacement for some nerve injured patients to communicate motor functions with a prosthesis device. Decoding algorithms translate ensemble of firing rates to the intended function. Firing rates for each individual neuron are obtained from labeling the detected spikes. This labeling process-also known as spike sorting-could be done from the range of fully automated to a heavily operator dependent manners. On the other hand we could use merits of both automation and operator's watch in a semi-supervised approach. In this study we explored the application of a self-training SVM classifier algorithm to label spikes with a small training dataset. Result shows the proved monotonically increasing convergence and consequently the ability of this algorithm to significantly reduce the operator's effort for continuous supervision. It provides in addition a significant improvement with respect to the previously used SVMs.