This paper describes a top-down word image generation model for holistic handwritten word recognition. To generate a word image, it uses likelihoods based, respectively, on a linguistic model, a segmentation model, and a character generation model. In the recognition process with respect to a given input image, it first generates, for each word in a dictionary of possible words, a word image that approximates as closely as possible the input image. The model next calculates distance values between each generated word image and the input image and selects for recognition that generated word image having the smallest distance value. The proposed method has been evaluated in an experiment using handwritten word images, and results show it to be effective for use in handwritten word image recognition.