In this paper, a memory efficient, statistical data driven approach is proposed and succesfully tested for grapheme-to-phoneme (G2P) conversion. In our system, a dynamic programming (DP) based fast algorithm is formulated to estimate the optimal joint segmentation between training sequences of graphemes and phonemes. A statistical language model is trained to model the contextual information between grapheme and phoneme segments. A two-stage fast decoding algorithm is also proposed to recognize the most-likely phoneme sequences given the input test word and the n-gram graphone models. Experimental results show that this system has similar recognition accuracy as a decision-tree based G2P system and requires much less memory and processing time.