This paper describes a new top-down word image generation model for word recognition. This model can generate a word image with a likelihood based on linguistic knowledge, segmentation and character image. In the recognition process, first, the model generates the word image which approximates an input image best for each of a dictionary of possible words. Next, the model calculates the distance value between the input image and each generated word image. Thus, the proposed method is a type of holistic word recognition method. The effectiveness of the proposed method was evaluated in an experiment using typewritten museum archive card images. The difference between a non-holistic method and the proposed method is shown by the evaluation. The small errors accumulate in non-holistic methods during the process carried out, because the non-holistic methods can't cover the whole word image but only part images extracted by segmentation, and the non-holistic method can't eliminate the black pixels intruding in the recognition window from neighboring characters. In the proposed method, we can expect that no such errors will accumulate. Results show that a recognition rate of 99.8% was obtained, compared with only 89.4% for a recently published comparator algorithm.