Extracting Unit Embeddings Using Sequence-To-Sequence Acoustic Models for Unit Selection Speech Synthesis.