Pronunciation of proper names is known to be a difficult problem, but one of great practical importance for both speech synthesis and speech recognition. Recently a few data-driven grapheme-to-phoneme conversion techniques have been proposed to tackle this problem. In this paper we apply the joint n-gram model for bi-directional grapheme- to-phoneme conversion, which has already been shown to achieve excellent results on general tasks, to the more specific task of converting between name pronunciations and spellings. The performance of our technique on generating name pronunciations exceeds that of other techniques even when they use additional information. We find the reverse task, of generating orthographic transcriptions from phonemic input, to be a much more difficult task for names than for common words. However, we derive valuable information from our results about the potential of sub-lexical recognition of novel proper names.