Restoring Lost Speech Components with Generative Adversarial Networks for Speech Communications in Adverse Conditions