Pronunciation Error Detection In Pakistani L2 Mandarin Based On Articulatory Features And Multitasking Learning

In recent years,with the development of the "Belt and Road" initiative in China,many Pakistanis have chosen Mandarin Chinese as their second language(L2),and mastering authentic pronunciation has become a difficult and important point for many learners.Computer-assisted pronunciation training(CAPT)can help L2 learners master correct pronunciation through pronunciation error diagnosis and feedback.However,most current pronunciation error detection systems are based on phonemes.Only errors that have a large difference from standard pronunciation can be recognized at the phoneme level.It cannot detect and correct pronunciation errors that deviate relative to standard phonemes,and cannot provide more accurate feedback on pronunciation errors.In response to the problems of existing pronunciation error detection systems,this paper constructs a pronunciation error detection system based on articulatory features as detection targets.It detects the state of the pronunciation organs of speech in pronunciation and compares the detection results with the standard phonetic pronunciation.Based on the identified pronunciation errors,feedback and correction opinions are given.Firstly,according to the characteristics of Mandarin Chinese pronunciation and the common problems in the pronunciation of Pakistanis learning Mandarin,four types of articulatory features,including Mandarin vowels and consonants,were designed for the recognition target,describing the pronunciation positions and methods of Mandarin and identifying errors in pronunciation.Secondly,this paper selects a articulatory features detection model that performs well in processing context information to detect continuous Mandarin speech.Therefore,a combined convolutional neural network(CNN)and bidirectional long short-term memory network(Bi LSTM)are used to construct the articulatory features detection model in this paper.The experimental results show that using the CNN-Bi LSTM network to detect articulatory features has an average accuracy rate 5.8% higher than using the long short-term memory network(LSTM).Thirdly,since the generation of speech involves the coordination of multiple pronunciation organs,there is a potential connection between each pronunciation organ and the pronunciation method.This connection can help identify articulatory features.Therefore,based on the high correlation between the recognition targets of articulatory features,this paper introduces the theory of multi-task learning to explore the relationship between multiple articulatory features recognition tasks and find the optimal multi-task articulatory features detection mode.The experimental results show that using multi-task learning for articulatory features detection has an average accuracy rate 3.0% higher than using single-task learning.Next,the optimized CNN-Bi LSTM model was applied to evaluate pronunciation errors made by Pakistanis learning Mandarin as their second language.This paper established a Mandarin L2 pronunciation corpus for Pakistanis and manually annotated and corrected the phonemes in the corpus to obtain the actual pronunciation situation in the corpus.Then the articulatory features detection model was used to detect the articulatory features of the L2 corpus,and the average accuracy rate reached 79.7%,proving that the modeln effectively detect L2 articulatory features.Error detection and statistics were performedon the L2 corpus,and examples of error detection were presented to analyze the problems in L2 pronunciation.Finally,the current status quo of pronunciation detection and correction in CAPT systems on the market was analyzed.Based on the constructed pronunciation detection and correction model,a Mandarin Chinese pronunciation error detection system based on articulatory features was built using mainstream Java Web technology to help L2 Mandarin learners with pronunciation training.