Self-repairs give us a key to understand how a human generates utterances. In this paper, we report a result of our analysis of timing factor of selfrepair. We categorize an error repaired by a stage it is generated, and investigate relation between a category and timing factor of the repair. The result shows that it takes more time to repair an error generated at earlier stage of generation. From the result, we develop a model of utterance generation. Finally, we try to account for some characteristics of self-repair with the model.