This paper discusses a method for training a garbage model for a keyword spotter. Conventionally, minimum error training of garbage models, which have been shown to be effective, required a large amount of keyword speech samples as training data. So, on changing a keyword set, a process of collecting a large amount of speech samples has been required. In order to solve this problem, this paper proposes a training method which needs no keyword speech samples, so a keyword set can be changed quickly and economically. Experimental results show the spotting performance can be improved by a training process without raw keyword speech samples.