| With the vigorous promotion of Mandarin,Mandarin has been widely learned as a second language.Because Tibetans have their own language and characters,Tibetan students are affected by the pronunciation of their native language.It is difficult to standardize the pronunciation of Mandarin,Tibetan students’ Mandarin pronunciation has obvious "national cavity and national tone",so it is particularly important to help Tibetan students learn Mandarin well.Computer assisted pronunciation training system(CAPT)can detect learners’ wrong pronunciation and provide detailed guidance feedback,which is helpful for second language learners to improve their pronunciation level more effectively.Mispronunciation detection and assisted learning are the core technologies of CAPT,and its research results are of important significance and practical value to supplement the existing learning resources and provide learners with barrier free and personalized learning methods.This thesis takes the pronunciation of Mandarin of Tibetan students as the research object,and takes the mispronunciation detection and assisted learning of tone,initial and final as the research content.Combined with deep learning technology and the knowledge of language phonetics,this thesis makes an in-depth study on the technology of mispronunciation detection and assisted learning of Mandarin of Tibetan students.The work and innovation of this thesis are as follows:1.Design and improve the structure of mispronunciation detection and assisted learning system.According to the process of artificial subjective detection and traditional language teaching,the system is divided into four modules: acoustic model training module,mispronunciation detection module,standard pronunciation transcription module,and assisted learning feedback module.In the feedback part of the system,on the basis of previous text-only feedback,the part of the assisted learning feedback module adds speech synthesis feedback.With the increase of corrective feedback forms,the system can not only provide learners with specific mispronunciation information,but also provide virtual teachers with teaching pronunciation,which can more effectively promote the learning efficiency of second language learners.2.Corpus design and establishment.Corpus design includes cross language training corpus design and mispronunciation test corpus design.According to the pronunciation characteristics and rules of Mandarin and Tibetan,this thesis establishes a cross language training corpus and a mispronunciation test corpus.The cross language training corpus is mainly used to train models.The mispronunciation test corpus is mainly used to test the performance of system mispronunciation detection.Then,according to the tagging rules proposed in this thesis,the manual tagging is carried out on the corpus.3.Research of mispronunciation detection task and assisted learning task of Tibetan students’ Mandarin.A mispronunciation detection model based on end-to-end CNN-DFSMN-CTC is built,and an assisted learning model based on end-to-end Tacotron speech synthesis is built,and an assisted learning model based on end-to-end Tacotron2-Wave RNN speech synthesis is built.Based on the above three models,the experiment of mispronunciation detection and the experiment of assisted learning of Tibetan students’ Mandarin is completed.The results show that the mispronunciation detection method based on CNN(6)-DFSMN(12)-CTC model has 89.34% detection accuracy,7.51% False Rejection Rate and 24.84% False Acceptance Rate.Compared with the previous tasks of Tibetan students’ Mandarin mispronunciation detection,the detection accuracy is improved,the detection performance is outstanding,and it is closer to manual detection.Based on the results of the experiment,the specific mispronunciation information of Tibetan students in three aspects: initial,final and tone is obtained.The mispronunciation of Tibetan students in Mandarin is analyzed from three perspectives: initial,final and tone.Tacotron2-Wave RNN-based speech synthesis assisted learning method with MOS value of 4.21 and MCD value of 9.483,based on the speech synthesis of Tacotron2-Wave RNN model is more suitable for assisted learning tasks of system. |
