Adversarial Learning of Intermediate Acoustic Feature for End-to-End Lightweight Text-to-Speech.