Research on Understanding of English Utterances by Second Language Speakers

在这项研究中，我们构建了一个学习者语料库，其中每个表达是由英语母语者主观评价，我们开发了一个基于对话的CALL系统，显示在日本的回答问题的系统，以限制他们的话语，使用学习者语料库。CALL系统不仅采用了使用ngram和有限状态自动机（FSA）的常用自动语音识别器，而且还采用了使用范围外话语模型的自动语音识别器，其中，所述自动语音识别器被设计用于接受目标域中的话语。在这项研究中，我们还提出了一种新的分类方法，分类的语音识别结果从识别器使用FSA接受或拒绝。该分类方法采用多个识别器的声学似然度之间的差异以及多个识别器的输出之间的编辑距离作为分类器的特征。

In this research we constructed a learner corpus in which each expression was subjectively evaluated by English natives, and we developed a dialogue-based CALL system which showed answers in Japanese to questions from the system to restrict their utterances, using the learner corpus. The CALL system adopts not only commonly used automatic speech recognizers using ngram and finite state automata (FSA) designed for accepting utterances in a target domain but also one using a model of out-of-scope utterances. In this research we also propose a novel classification method which classifies the speech recognition result from the recognizer using the FSA into accepted or rejected. The classification method adopts differences among acoustic likelihoods from the multiple recognizers as well as edit distances among the multiple recognizers’ outputs as features for the classifier.