Research On Variational Mode Decomposition And Specific Instruction Extraction Of Vehiclemounted Speech

IFLYTEK,Baidu,Alibaba,Tencent and other companies have conducted a lot of research on vehicle-mounted speech signals,and applied them to the analysis and recognition of vehiclemounted speech,such as vehicle navigation,vehicular telephone,etc.Due to the complex environment in the vehicle,the application of speech signals is greatly limited by the nonstationary characteristics of noise.In this paper,Variational Mode Decomposition(VMD)algorithm is used to decompose the collected signals and extract specific speech signals.The main research of this paper is as follows:1.In the existing related literature of CNKI,VMD algorithm is not used to process vehiclemounted speech signal.This paper first studies VMD algorithm,applies the calculation center frequency and instantaneous frequency mean method to VMD algorithm,and determines the most appropriate decomposition layer,which solves the problem that the decomposition layer number of VMD algorithm is difficult to determine.To determine the number of layers decomposed by the VMD algorithm,the first step is to pre-decompose the signal to obtain multilayer components,and then determine the number of layers decomposed by judging whether the center frequency is close to.Then,the instantaneous frequency mean method was used to determine whether the mean value decreased significantly to obtain the decomposition layer number of VMD and reduce the calculation amount.2.After the decomposition of VMD,the multi-scale permutation entropy of the Inherent Mode Function(IMF)obtained is calculated,and the IMF is classified and recognized by support vector machine to extract specific speech signals.In practice,VMD decomposition is first carried out for specific speech signals,and then each component is classified and recognized.The overall accuracy rate reaches 98.33%,and the classification and recognition of speech signals reaches 100%.3.Calculate the characteristic quantities of specific speech signals,such as Mel Frequency Cepstrum Coefficient(MFCC),establish Hidden Markov Model(HMM),use the K-means method to cluster the features,and complete the recognition of the extracted speech through training and testing.4.Particle Swarm Optimization(PSO)algorithm and Krill Herd(KH)algorithm were used to optimize the decomposition layer number and penalty factor of VMD,and the results of optimal parameters were obtained.Since only the number of decomposition layers is considered when the VMD algorithm is used for decomposition,one of the optimal solutions can be obtained.However,the PSO-VMD algorithm and KH-VMD algorithm can obtain better extraction results by considering both the number of decomposition layers and the penalty factor.5.A large number of experiments have been done by using the collected vehicle-mounted speech signal data.When the car is running normally,the recorded alarm is used as the interference noise,and at the same time,the person sends out instructions speech to get the collected signal.The experimental results show that the speech recognition accuracy of the undecomposable vehicle-mounted speech signal is 92.5%,the accuracy of VMD decomposition is 97.2%,the accuracy of VMD algorithm optimized by particle swarm optimization is 99.2%,and the accuracy of VMD algorithm optimized by krill algorithm is 98%.It can be seen that the VMD algorithm and the optimized VMD algorithm have made some progress in the extraction of vehicle-specific speech signal results.