Learning Word Embeddings: Unsupervised Methods for Fixed-size Representations of Variable-length Speech Segments.