DNN-Based Talking Movie Generation Using Minimum Generation Error Criterion and Adversarial Learning with Face Direction Consideration