EEG is an extensively used powerful tool for brain computer interface due to its good temporal resolution and ease of use. The signals captured by multichannel EEG recordings contribute to huge data and often lead to the high computational burden on the computer. An optimal number of electrodes that capture brain signals relevant to the purpose can be used, excluding the redundant and non-contributing electrodes. In this study, we propose an optimization technique on common spatial pattern for channel selection. The implementation of optimization is done as a sequential quadratic programming problem of fast convergence. Extensive experimentation is done to show that the proposed method induces large variance between two tasks of brain action related to sub vocalized speech.