Based on the framework of error-correcting output coding (ECOC), we formerly proposed a multi-class classification method in which mis-classification of each binary classifier is regarded as a bit inversion error based on a probabilistic model of the noisy channel. In this article, we propose a modification of the method, based on localized likelihood, to deal with the discrepancy of metric between assumed by binary classifiers and underlying the dataset. Experiments using a synthetic dataset are performed, and we observe the improvement by the localized method.