Raters’ constant errors, represented mostly as leniency errors and halo effect, have been examined for the L2 Korean speaking assessment by native Korean speakers. Fifteen Korean raters assessed twenty-seven native English, Japanese, and Chinese learners of Korean with three levels of beginning, intermediate, and advanced groups. FACETS program with multi-facets Rasch model has been used to analyze the data. The results are as follows: (1) test-takers’ mother tongue has little impact on Korean raters’ rating patterns in a whole, but they tend to grade leniently to English learners of Korean, (2) test-takers’ levelling by L2 Korean proficiency show statically significant influence on raters’ ratings. The study finds that Korean raters could have meaningfully constant errors on their assessment, even though they don’t agree with committing constant errors in the survey. In details, they have more leniency grading to English learners of Korean, while they have halo effects to the beginning stages of L2 Korean learners. The result implies that Korean raters should carry out the authentic speaking assessment training before the real tests.(Konkuk University)