Data-Driven Learning for Pronunciation