An Empirical Study of Visual Features for DNN Based Audio-Visual Speech Enhancement in Multi-Talker Environments