In recent years, the signal processing opportunities with the multi-channel recording and the high precision detection provided by the development of new extracellular multielectrodes are increasing. Hence, designing new spike sorting algorithms are both attractive and challenging. These algorithms are used to distinguish the individual neurons' activity from the dense and simultaneously recorded neural action potentials with high accuracy. However, since the overlapping phenomenon often inevitably arises in the recorded data, they are not accurate enough in practical situations especially when the noise level is high. In this paper, a spike feature extraction method based on the Wavelet Packets Decomposition and the Mutual Information is proposed. This is a highly accurate semi-supervised solution with a short training phase for performing the automation of the spike sorting framework. Further, the evaluations are performed on different public datasets. The raw data is not only suffered from multiple noises (from 5% level to 20% level) but also includes various degrees of overlapping spikes at different times. The clustering results demonstrate the effectiveness of our proposed algorithm. Also, it achieves a good anti-noise performance with ensuring a high clustering accuracy (up to 99.76%) compared with the state of art methods.