Research On The Automatic Evaluation Model Of Spoken Language Integrating Non-native Chinese Speakers

As a large multi-ethnic country,in people's daily communication,in addition to Mandarin,which is widely used as a standard universal language,there are also various dialects and minority languages due to regional cultural differences.In previous studies,a scoring model was constructed through automatic evaluation of spoken Chinese in reading questions,and emotional evaluation indicators were innovatively incorporated.On this basis,this research collects speech data from elementary school students in Suzhou and Xinjiang.Students in Xinjiang,as non-native Chinese learners,can expand the sample size of the automatic scoring model and improve the applicability of the model.At the same time,in the process of constructing the model,advanced speech processing technology was used to analyze the pronunciation characteristics of students in the two regions,and to give targeted suggestions for students in Xinjiang to learn Mandarin Chinese in order to improve the teaching of Mandarin Chinese in Xinjiang Provide some new ideas.In the speech corpus selection and speech feature analysis in this study,the previous three dimensions of ancient poem reading,accuracy evaluation,fluency evaluation,and sentiment evaluation are still used.At the same time,three expert teachers were invited to score the voice data according to certain scoring standards,and finally the expert scores were used as the standard scores for model training and construction.Among them,the accuracy evaluation uses the full-dimension voice evaluation API interface of the i Flytek open platform,and extracts the five scores of accuracy score,overall impression score,completeness score,phonological score,and tonal score,which are included in different groups The scoring model is finally determined to be the final accuracy evaluation score based on the four average scores of phonological score,tonal score,completeness score and overall impression score.In terms of fluency evaluation,short-term energy and endpoint detection technology are used to remove silent segments and extract the number of pauses,and compare them with standard speech to calculate the speech rate and pause score.Emotion evaluation uses open source Python code to extract features,calculate the cosine similarity with standard speech,and obtain the final emotional score.Using multiple linear regression,nonlinear regression and DNN classifier training scoring model and other construction methods,with accuracy,sentiment and fluency scores as independent variables,manual expert scores as dependent variables,to build an automatic scoring model.Finally,the correlation test was performed on the model prediction score and the expert score,and the results showed that the correlation coefficients of the multiple linear regression model and the nonlinear regression model both exceeded 90%.Compared with the multiple linear regression model obtained in the previous study,it is found that the accuracy of the voice evaluation of samples in different regions has been improved,indicating that the effects of the multiple linear regression and non-linear scoring models in this study have reached expectations.The effect deviation of the DNN classifier model may be related to the small sample size.In future studies,the sample size can be further expanded to try.In the analysis of the difference between the speech data of students in Xinjiang and Suzhou,in addition to comparing the accuracy and fluency in the evaluation dimension,the emotional dimension of the speech data was also compared in detail.The emotion of speech is mapped in the data features,often reflected in energy,pitch,sound intensity,and formant.Through data analysis and comparison,it is found that in Xinjiang students in the process of reading Mandarin,Yangping is easy to be confused as a missing tone,the upper tone is not obvious and other errors,and the light and accent in the reading is not handled properly.Therefore,teachers can carry out targeted teaching according to the rule of error to improve the standard of Mandarin Chinese of Xinjiang students.