In this experiment, we are looking for a non-native learner's error to have a score for a given target phoneme in a given context, that is significantly distant from the scores that a native model speaker has. As task data, the continuous speech, which may contain prosodic transfer, was chosen. In general, learners may transfer the prosodic patterns of their native languages (L1) into the target language (L2). Therefore, we investigate how relative threshold functions will work in this task. For this purpose, we prepared a total of 13 sentences from Japanese textbook. To automatically detect pronunciation errors, the relative thresholds based on a native model speaker were used. Due to experimental limitations, we have not investigated the contextual effects for specific speech materials. Based on the score of a native model speaker, we investigate how effectively it can detect pronunciation errors.