Research On Automatic Pronunciation Detection System Based On Machine Learning

With the rise of online learning,computer assisted language learning has become the choice of more and more language learners.As the two cores of computer-assisted language learning,pronunciation error detection and diagnostic feedback can analyze the learner's pronunciation problems and give pronunciation correction opinions,thus improving learners' pronunciation level and learning efficiency.Compared with classroom language teaching subject to time and space constraints,computer automatic pronunciation detection has many advantages such as real-time,convenient and efficient.At present,most academic research on pronunciation error detection only focuses on the detection of pronunciation errors,and ignores the importance of feedback correction.In order to provide learners with more intuitive pronunciation correction Suggestions,this paper studies the types of phoneme level pronunciation errors caused by learners' non-standard pronunciation actions,Combining machine learning algorithms for classification,error detection experiments on six types of errors: Rising,Lowing,Fronting,Backing,Lengthing,and Shorting.And using the Internet technology to design and test the pronunciation error detection model in the Web system integration,the automatic pronunciation correction system was developed.Firstly,after evaluating the acoustic features and the phonetic corpus,the paper proposes a pronunciation classification and error detection model based on MFCC-RF.Constructing a classification error detection model using the extracted 39-dimensional Mel Frequency Cepstral Coefficents(MFCC)acoustic feature as input to a random forest(RF)classifier.By analyzing the experimental results,the MFCC-RF mispronunciation model has achieved high mispronunciation accuracy in Raising,Lowing and Shorting.However,because the sample data of mispronunciation types are not evenly distributed,the sample data of the other three types of mispronunciation types are relatively small,Therefore,the pronunciation classification error detection model based on MFCC-RF is only applicable to Raising,Lowing and Shorting three types of error detection,with a small range of error detection.Deep learning has proven to be very suitable for pattern recognition and extraction of complex features in recent years.In order to expand the range of pronunciation error detection types on the basis of the mispronunciation classification model based on MFCC-RF,and further improve the mispronunciation accuracy.Extract deep hidden information contained in acoustic features through deep neural network as input to machine learning classification algorithm.A classification error detection model based on DBN-SVM is proposed,and the OneClass idea of support vector machine is used to solve the problem of sample data imbalance.The pronunciation classification error model based on dbn-svm added the detection of three error types,including Fronting,Backing and Lengthing,and completed the classification error detection of all six types of pronunciation error,and verified the effectiveness of the model through experiments.Then,combined with the popular trend of the Internet electronic teaching method,this paper USES the Java Spring framework combined with the webpage related technology to carry out the preliminary design and development of the Web system for the pronunciation error model.It is a platform for learners to correct and improve pronunciation freely and online,laying a certain foundation for the future development of mobile online English pronunciation learning.Finally,the paper summarizes and forecasts the full text.This paper expounds the work to be solved and perfected in the construction of the automatic pronunciation error detection system and the next step.