A major source of recognition errors, out-of-vocabulary (OOV) words are also semantically important; recognizing them is, therefore, crucial for understanding. Success, so far, has been modest, even on very constrained tasks. In this paper we present a new approach to unlimited vocabulary speech recognition based on using grapheme-to-phoneme correspondences for sub-lexical modeling of OOV words, and also some very encouraging results we obtained with our approach on a large vocabulary speech recognition task.