英语发音自动评测中单词删除错误的检测方法

A Word-deletion Error Detection Method for English Automatic Pronunciation Assessment

基于HMM后验概率分数的自动发音评测方法中,强制匹配获得语音的音素级切分信息是重要的一步。切分结果是否准确直接影响到后验概率分数的可靠性。通过对单词发音网络增加可跨越分支,实现单删除错误检测功能,可降低说话人删除错误对前后单词切分准确性的影响。实验结果显示,单词删除错误检出率达到87.8%。

To obtain phonetic segmentation information by force alignment is an important step in HMM-based automatic pronunciation assessment. The accuracy of segmentation has a direct impact on the result of posterior probability score. Add a parallel path span word recognition network can reduce the segmentation errors of adjacent words caused by word-deletion pronunciation errors. The experimental results indicate that the accuracy of worddeletion errors detection reaches 87.8%.