Sparse Representation with Optimized Learned Dictionary for Robust Voice Activity Detection