English mispronunciation detection based on improved GOP methods for Chinese students

In this paper, we proposed two approaches to detect mispronunciation in spoken English for Chinese Students which is based on improved Goodness Of Pronunciation (GOP) algorithm. We adopted a modified Maximum Likelihood Linear Regression (MLLR) to adjust the acoustic model which can reduce the mismatch between native original model and adaptive data from non-native speakers. Then we could calculate the ameliorated GOP value to improve the performance of phone-level pronunciation error detection. Besides, as Chinese students are likely to be influenced by their mother tongue in their oral English training, we collected the common pronunciation error patterns of Chinese students by introducing priori linguistic knowledge and established a phonemes set that are easy to confuse for optimizing the GOP probability space. The mispronunciation detection system with the above ways could review input speech and detect the flawed phone to allow the non-native learners to correct the mispronunciation duly. The experimental results suggested that the modified GOP method reached good effect of English pronunciation error detection for Chinese students.