Improve emotional speech synthesis quality by learning explicit and implicit representations with semi-supervised training.