Subjective intelligibility of deep neural network-based speech enhancement