Exploiting Nonnegative Matrix Factorization with Mixed Group Sparsity Constraint to Separate Speech Signal from Single-channel Mixture with Unknown Ambient Noise