We present a novel phoneme-based student model for spelling training. Our model is data driven, adapts to the user and provides information for, e.g., optimal word selection. We describe spelling errors using a set of features accounting for phonemic, capitalization, typo, and other error categories. We compute the influence of individual features on the error expectation values based on previous input data using Poisson regression. This enables us to predict error expectation values and to classify errors probabilistically. Our model is generic and can be utilized within any intelligent language learning environment.