Research On Speech Enhancement Algorithm Based On Variational Mode Decomposition

With the rapid development of artificial intelligence,such as human-computer interaction,speech recognition and so on,which need speech participation,are inseparable from the front-end technology of speech enhancement.In addition,the core content of speech enhancement needs to be discussed,which is how to improve the quality of speech in complex noise scenes while effectively removing noise,so as to conform to the auditory characteristics of human ears.The existing speech enhancement algorithms can achieve good results under weak background noise.However,with the change of noise environment,such as the switching of complex scenes and the sharp decline of input SNR,it is of great significance to make the speech enhancement algorithm balance the denoising effect and speech quality.As an effective time-frequency analysis method,the variational mode decomposition(VMD)algorithm is based on the characteristics of the signal itself,so it has the adaptability in signal decomposition.Therefore,VMD has great advantages in dealing with non-stationary and non-linear signals,and speech signals are the most representative.Taking VMD algorithm as the core,this paper carried out in-depth research in the field of speech enhancement,and improved some problems existing in VMD algorithm,so as to design an effective speech enhancement method.The main work is as follows:1.Multi resolution analysis is used to solve the problem of fuzzy mode number in VMD.Firstly,according to the characteristics of human ear in multi-resolution analysis,the speech signal is analyzed by multi-resolution analysis.Combined with a large number of experiments and analysis,it is found that the optimal decomposition level is 6-7 layers.Then,the speech signal is decomposed by wavelet,and the Spearman correlation coefficient between the high and low frequency sub signals in each scale is calculated,which is used as the basis to determine the number of center frequencies of the speech signal,so as to solve the problem of Fuzzy Mode number in VMD.2.The number of the interrelation between the estimated noise and the variational mode function is taken as the basis for mode selection,and the precise denoising of the speech signal is realized.In order to get the estimated noise,the first three low-frequency sub signals in the multi-resolution analysis results are reconstructed.The cross-correlation coefficient between each variational mode function and the estimated noise is calculated,and the speech mode function is accurately selected by setting the adaptive threshold.By comparing the speech signal enhanced by MEMD-VMD algorithm,it is clearly proved that the method of accurately selecting the mode function can eliminate the noise to the greatest extent and preserve the speech information at the same time.3.On the basis of noise elimination,the speech details are retained to further improve the speech quality.In order to preserve speech details,the VMD decomposition of the existence of small amplitude and slow transformation of the variational mode function as a part of speech reconstruction,in order to achieve the purpose of improving the quality of speech.The experimental results show that the reconstructed speech signal not only performs well in denoising,but also improves the speech quality.