Adaptive Listening Difficulty Detection for L2 Learners Through Moderating ASR Resources

Teaching listening skills to those learning a second language (L2) is one of the most challenging tasks mainly because predicting L2 listening difficulties is not always straightforward. Complex processes are involved in decoding connected speech, constructing meaning, and comprehending the audio material. Many studies have attempted to identify the significant factors leading to listening difficulties, yet, a comprehensive model is to be constructed. We argue that an automatic speech recognition (ASR) system with limited training can be viewed as a rough model for an L2 listener with particular language proficiency. We proposed a method to select the training samples for the ASR system to match the mistakes of L2 listeners when listening to the authentic listening materials. This model can predict the learners’ listening difficulties, thus allowing for generating tailored captions to assist them with L2 listening.