In this study, we propose a new application using multichannel source separation technique based on Non-negative Tensor Factorization which is generalization of Non-negative Matrix Factorization. The application enables us to extract sources in certain direction that the user specified and to manipulate as the user wishes to be such as zoom up or zoom down. All the components of factorization are supposed to be trained such that each component should be assigned to diverse direction so that the system can simply choose and extract only the components assigned to the direction of user's demand. However, due to the high possibility of local minimum, it is likely that the sources supposed to be trained in the certain direction are wrongly trained in the different directions. To overcome this problem, we propose a new initialization method based on Inter-channel Intensity Difference between the channels. This improves the quality of separation measured by SDR, SIR and SAR, which are standard metrics of source separation technique nowadays.