Rethinking complex-valued deep neural networks for monaural speech enhancement