Study On Method Of Mask Speech Correction

In the case of underwater communication,high-altitude missions and rescue operations,which require humans to wear full masks,the closure of the full masks make the speech distortion.The distortion hampers the normal communication of human beings,and makes human activities a variety of inconvenience.Therefore,it is of great significance to study on mask speech correction for the scientific exploration,national defense construction,social life,rescue operations and military activities.Mask speech correction is a nonlinear modeling problem,which is modeling the feature parameters of the mask speech and normal speech by the nonlinear mapping model,so as to achieve the effect of correcting the speech.In order to study the application of nonlinear mapping ability of neural network in the mask speech correction system,related works are as follows:(1)Analysis of speech characteristics of mask speech.Based on the analysis of speech signal,the characteristics of the cover speech are analyzed in time and frequency domain.And the characteristics of the mask speech are summarized.(2)Research on correction model based on GRNN and LSP parameters.Using GRNN to model the correction models for LSP parameters with the vocal tract information and the prediction error signals with the excitation sequence information.On this basis,k fold cross validation method and Particle Swarm Optimization(PSO)algorithm are used to optimize the GRNN parameter,which can improve the prediction accuracy.The experiment results show that the two algorithms both can optimize the parameters of the network,and the correction performance of GRNN optimized by PSO is better.(3)Research on correction model based on GRNN and STRAIGHT spectrum parameters.Owing to the discontinuity problem caused by the method of correcting LPC model based on GRNN,STRAIGHT model is used to extract the independent pitch period and the spectral parameters.Then using GRNN to model the correction model with spectral parameters.The experiment results show that the method makes the discontinuous spectrum smoothing,and improves the spectrum similarity of corrected speech and normal speech.(4)Statistical mean of training samples based on K-means algorithm.In order to ensure the accuracy of prediction and speed up the training speed of the correction model,the K-means algorithm is used to cluster the training samples and find the center values.The experimental results show that this method can not only shorten the training time of the correction model,but also improve the generalization ability of the mask speech correction system in the case of a certain number of training samples.