Research On The Pronunciation Error Detection Of Tibetan Students' The National Common Language Based On CNN

In today's society with great development and integration,people need to be able to speak the National Common Language so that both sides can understand it accurately.Therefore,improving the pronunciation level of the National Common Language has become the requirement of personal development.In recent years,with the implementation of China's western development plan and the deepening of economic and cultural exchanges with the mainland,the Tibetan economy has developed rapidly.In this context,the Tibetan people's demand for learning the National Common Language is growing.It is particularly important to help Tibetan speakers learn the National Common Language effectively.Therefore,it is necessary to make a deep and systematic study on the automatic detection technology of pronunciation errors for Tibetan speakers.At the same time,the automatic detection of pronunciation errors is the core technology of computer assisted pronunciation training system(CAPT).The research results are of great significance and practical value to supplement the existing learning resources and provide learners with barrier free,anxiety reducing and personalized learning methods.This thesis takes the pronunciation of the National Common Language of Tibetan native speakers as the research object,and takes the pronunciation errors of initials,finals and tones as the research content.Combined with the knowledge of language phonetics and deep learning technology,this thesis makes an in-depth study on the automatic detection technology of pronunciation errors of the National Common Language of Tibetan native speakers.The main work and innovation are as followsThe design and establishment of corpus.The PSC-test-1 speech database is designed according to the pronunciation characteristics and rules of Chinese Tibetan bilingualism.There are 2268 sentences and 10 speakers in the database.The database is used to test the performance of the system.The PSC-train-1 data set was selected from four open source national standard pronunciation speech databases for acoustic model training.Corpus data annotation.The corpus used in this thesis adopts the extended vowel annotation method,so that each syllable has a strict "sound rhyme tone" ternary structure.Therefore,seven zero initials are extended.According to this ternary structure,pronunciation errors are divided into three categories: initial pronunciation errors,vowel pronunciation errors and tone pronunciation errors.According to their respective structural characteristics,these three types of errors are subdivided For 64 kinds of pronunciation errors.Pronunciation error detection of the National Common Language.Combined with Connectionist Temporal Classification and neural network technology,an end-toend pronunciation error detection model is built,and under the model,the pronunciation error detection experiment of Tibetan students speaking the National Common Language is completed.The experimental results show that the detection accuracy rate is 88.35%and the combined error rate is 14.91%.It can effectively detect the pronunciation errors of Tibetan students when they speak the National Common Language from the angles of initials,finals and tones.