TAM-BLSTM Based Continuous Mandarin Tone Error Detection And Analysis For Pakistani

With the development of China's "the Belt and Road",many Pakistanis learn Mandarin as their second language(L2).The speech error detection module of computer assisted pronunciation training(CAPT)is very helpful for L2 learners to improve their mastery of the language.For Mandarin,the pronunciation of tone is particularly important,and it is particularly difficult for Pakistanis whose mother tongue is a non-tone language.However,at present,there is less research on the tone of Pakistani Chinese learners in China.Therefore,this paper focuses on the error detection of tone.In order to effectively help L2 learners master the correct tone in their daily communication and dialogue in CAPT,the research of tone recognition and error detection of continuous speech is very important.This paper studies the tone error detection of L2 Mandarin continuous speech from two aspects: feature extraction and classification algorithm,in order to build a tone error detection model of Mandarin continuous speech flow with high error detection rate,so as to provide learning suggestions for L2 learners' Mandarin tone errors.Firstly,in view of the complex tone changes in continuous speech and the characteristics of L2 learners' tone pronunciation,this paper makes corresponding improvements to the fundamental frequency generation model,namely the Target Approximation Model(TAM),and proposes a TAM parameter extraction model based on prosodic words.That is,taking the prosodic word as the unit,simulate the new fundamental frequency(F0)contour from the original F0 contour and generate the pitch target,and then export the relevant parameters of the pitch target,that is,TAM parameters,as the input features of the tone detection model.In the experiment,it is found that the accuracy of tone recognition is 2.3% higher when TAM parameter features are combined with fundamental frequency statistical parameters as the feature input of the model.Moreover,TAM parameters extracted based on prosodic words can better reflect the tone trend in continuous speech flow than those extracted based on syllables.The accuracy of using the former to train the model is 2.0% higher than that of the latter.Secondly,because the research object of this paper is mandarin continuous speech,the selected tone detection model needs to have the ability to process context information.Therefore,this paper selects a Bidirectional Long Short-Term Memory(BLSTM)network that can process forward and reverse time information to recognize and detect tones.From the experimental results,the accuracy of tone detection using BLSTM network is 3.8% higher than that using Long ShortTerm Memory(LSTM).Finally,the tone detection model based on TAM-BLSTM is applied to the tone detection of Pakistani L2 Mandarin corpus,and the accuracy rate is 78.3%,further,the pronunciation error detection is carried out according to the standard pronunciation,and the false acceptance rate and false rejection rate are 15.5% and 22.0% respectively.At the same time,according to the detection results,the tone pronunciation characteristics of learners in the corpus are analyzed and learning suggestions are given.