A Sparsity Based Preprocessing for Noise Robust Speech Recognition