Research On Scoring System Of Open-Ended Speech Based On Neural Network

With the continuous improvement of education informatization,the computer-assisted language learning(CALL)system is widely used in language teaching.In the spoken English tests of domestic universities,teachers need to manually score the spoken recordings of a large number of students.This is a repetitive and time-consuming task.Using the CALL system to automate this process will reduce the workload of teachers.At present,such a system has been successfully applied in read aloud task.However,the automatic scoring of open-ended speech is still the focus of research.Therefore,designing and implementing an intelligent scoring system for open-ended speech has important research significance and application value.This article combines deep learning technology and object-oriented design ideas to design and implement the scoring system.The system will score the spoken audio and spoken content separately through two scoring models,and add the two scoring results as the final score.The spoken content is obtained from the transcription of recording by an external speech recognition engine.This paper builds two types of scoring models based on different neural networks.In the first type of model,both the speech scoring model and the text scoring model are constructed using BP neural networks.The input features of these models need to be manually selected.This article has extracted seven types of features:pronunciation quality,fluency,content richness,topic relevance,grammar,vocabulary richness,and sentence structure.In the second type of model,both two scoring models are constructed using one-dimensional CNN and LSTM networks.This type of "end-to-end" scoring model does not require feature engineering.This article converts spoken recordings and spoken content into MFCC vectors and word embedding vectors,Then they are used as the input of the scoring model.Finally,the model is trained and tested by the research data includes 650 oral recordings from the tests of the Situational English Course at Beijing University of Posts and Telecommunications as well as their corresponding scoring data labeled by teachers.The experimental results show that the BP network model obtains better overall scoring performance when the training data set is relatively small.