Research On Recognition Of Several Typical Musical Instructments Based On Acoustic Features

In recent years,with the rapid development of computer science information technology,research on the acoustic aspects of audio signal have gradually become the research hotspot in music field.Musical instrument recognition based on acoustic features is the focus of the current music signal analysis,in order to improve the accuracy of instrument audio signals' recognition,this thesis mainly studied the collected audio of the six kinds musical instrument that were the guzheng,guitar,piano,accordion,harmonica and suona,this study mainly included the following several aspects:Firstly,aiming at the problem of low recognition rate of musical instrumental audio signals caused by additional noise in musical instrumental audio signals,the improved variational modal decomposition(VMD)was used to denoise the audio signals of musical instruments in this paper.In this paper,the audio signals of the instrument were decomposed into a series of stable narrow-band components by VMD firstly.Then through theway of the IMFs with the correlation coefficient greater than or equal to 0.5 and the effective information obtained by the wavelet threshold denoised the IMFs with correlation coefficient less than 0.5 were used to reconstruct the signal to improve the VMD.This paper studies the advantages and disadvantages of the acoustic signal denoising by wavelet threshold de-noising,empirical mode decomposition,VMD and improved VMD,the simulation results showed that the improved VMD algorithm is better than other de-noising algorithms in this paper.Secondly,in order to improve the classification accuracy after denoising,the sound features which fully reflect the sound characteristics are extracted from the musical instrument audio signal.The sound feature is a 24-dimensional combined feature of Mel frequency cepstrum coefficient and a first-order differential Mel frequency,based on the improved kernel principal component analysis(KPCA)in this paper.24-dimensional Mel frequency cepstrum coefficient and 24-dimensional a first-order differential Mel frequency were extracted as input parameters of the classifier.The experimental results showed that the 48-dimensional combined features of 24-dimensional Mel irequency cepstrum coefficient and 24-dimensional a first-order differential Mel frequency got a higher recognition rate than 24-dimensional Mel frequency cepstrum coefficient or 24-dimensional a first-order differential Mel frequency.Because of the dimensions of this combined feature were very high,in order to improve the speed of operation and the accuracy of classification recognition and reduce the amount of computation,the improved KPCA algorithm was used to reduce the dimensions of this combined features.The simulation results showed that improved KPCA removed more interference and retained more elements that reflect the timbre character of the audio signal than KPCA and PCA.Finally,the improved PSO optimized SVM is used as the classifier of musical instrument's audio signal.Due to the small sample number of experimental data,the support vector machine(SVM)was determined to used for classification recognition of the de-noising instruments' audion signal.The selection of SVM parameters has great influence on the recognition results.In the next simulation experiment,particle swarm optimization(PSO)and improved PSO were used to optimize the SVM's parameters.The improved PSO algorithm is an improvement on the adaptive inertia weight and asynchronous adaptive learning factor of PSO.The experimental results showed that SVM whose parameters were optimized by the improved PSO get higher recognition rate of musical instrument audio signals.