Developing Pronunciation Models in New Languages Faster by Exploiting Common Grapheme-to-Phoneme Correspondences Across Languages.