In language learners' speech, mispronounced words, word fragments, repairs, filled pauses, etc are often found, and they can be detected with ASR-based CALL systems. When learners are listening, some segments in a given utterance are often difficult to identify or misidentified due to lack of listening skill. In this study, we aim at detecting learners' listening breakdown to measure their listening skill. Listening skill is often quantified by imposing manual dictation on learners, but it has inevitable problems because manual dictation is generally an offline task. To solve the problems, oral dictation is imposed instead, and speaking breakdown is detected in the dictation utterances. Here, we assume that learners' speaking breakdown is attributed to their listening breakdown. This method is applied to measure their listening skill and to model its improvement exclusively through shadowing, which is oral dictation with a short delay and was introduced to language education originally as listening training. 35 Japanese university students attended a 42-day intensive shadowing training, and their shadowing utterances were analyzed to detect listening breakdown. Our model exhibits very monotonous improvement of listening skill as a function of how many days learners attended shadowing.