Research On Modeling Methods Of Vehicular Answering Speech Training And Evaluation System

Military parade is one of the major ceremonies to demonstrate national military might.There are strict training requirements and evaluation standards for the posture,salute,and answering speech of members in parade teams.In order to improve the training efficiency and ensure the fairness of evaluation,it is critical to establish a scientific and standard evaluation system during parade training.However,China still mainly adopts the traditional manual training and evaluation method at present,which is time consuming and strenuous.Besides,the evaluation results are likely to be affected by subjective factors.Therefore,there is a realistic need to develop a rational and efficient training and evaluation system to realize electronic and intelligent parade training and evaluation.Among various parade teams,members of the radar warning team usually need to be trained and evaluated in armoured cars.In this paper,answering speech samples from the radar warning team members such as ‘Good chairman' and ‘Serve the people' are chosen as the research object.Modeling methods of vehicular answering speech training and evaluation system are studied based on three evaluation indicators: volume,syllable and tone.The main research contents and results are as follow:(1)The answering speech samples are preprocessed and the data features are extracted.First,pre-emphasizing,adding windows,framing and detecting endpoint are used to preprocess answering speech signal.Then,some of speech feature parameters such as short-time energy,cepstrum and so on are extracted.Features of evaluation indicators such as volume value,syllable length,pitch and so on are calculated.And input features of regression prediction models are designed based on the feature data.Moreover,an improved pitch detection algorithm based on nonlinear combining function is proposed.The simulation results show that the improved method can effectively increase the accuracy of pitch detection compared with some traditional methods such as autocorrelation algorithm,cepstrum algorithm and so on.(2)An answering speech objective evaluation model based on evaluation indicator similarity intercomparison is established combined with the traditional evaluation method.By means of establishing the linear functional relationship between the features of evaluation indicators and corresponding scores,the answering speech to be evaluated and the standard answering speech are intercompared based on data similarity.Furthermore,final evaluation results are obtained according to different weight values of the evaluation indicators.(3)An answering speech score prediction model based on improved Stacking integrated learning is established combined with machine learning methods.First,in order to improve generalization of regression model,the Stacking integrated learning model,which uses support vector regression,rand forest and radial basis function neural network as the base learners and K-nearest neighbor regression as the meta-learner,is proposed.Then,in order to further improve prediction accuracy,the Stacking integrated model is optimized.The improved Stacking integrated model,which uses output feedback error of the base learners to weight the meta-learner input,is proposed.The experimental results show that the prediction accuracy of the Stacking integrated fusion model proposed in this paper is higher than the single basic model,and that of the improved Stacking integrated model attains further enhancement.In addition,the prediction results of the improved Stacking integrated model are in better agreement with practical manual evaluation scores,compared with the evaluation model based on similarity intercomparison.The experimental results prove that the improved evaluation model based on machine learning methods is feasible and effective.The research results are of guiding significance for developing a vehicular answering speech training and evaluation system.