This paper describes an adaptive, partial-word-level, writer-dependent, handwriting recognition system that utilizes the character n-gram statistics of the English language. The system exploits the linguistic property that very few pairs of English words share exactly the same set of character bigrams. This property is used to bring linguistic context to the recognition stage. The recognition is based on, estimating the probability of bigram co-occurrences between words. Preliminary experiments using naive features and limited training sets show that the system can recognize over 60% of words it has never seen before in handwritten form. The system has only few trainable parameters. In addition, incremental training is computationally inexpensive.