The study of L2 mispronunciation detection based on Mandarin landmarks

Abstract

It has been proved that phonetics knowledge or data-driven method based acoustic landmarks are useful in detecting mispronunciation. The acoustic landmarks obtained by the two methods are not completely consistent. It shall be studied which method is better. The role that the acoustic landmarks play in the mispronunciation detection task needs to be further explored. This paper compared the consistency of different acoustic landmark detection methods. The paper also compared the effect of mispronunciation detection through TNDD-GOP architecture and hybrid CTC/Attention model. The paper verifies the role of acoustic landmark method in mispronunciation detection task with weighted method. The experimental results show that the higher the weight of acoustic landmark is added the better the detection performance is got. The results show that the mispronunciation detection performance of the updated acoustic landmark based on phonetics knowledge is better than that of the data-driven method. The DA and FRR are improved by 3.38% and 1.38%.