A voice activity detection in mobile environments is not performed well due to arbitrary noises. In this paper, a robust voice activity detection framework for mobile devices is proposed. The unsupervised clustering and discriminative weight training of each cluster is employed to model various characteristics of arbitrary noises.