Speech Dereverberation Based On Elastic Network Regularization For Multi-Channel Sparse Priors

In places where the acoustic environment is closed,such as indoors and cars,the speech signals collected by the microphone will contain reverberation.As an important part of speech enhancement,the speech dereverberation algorithm can preprocess speech recognition and other technologies to dereverberation in the speech signal and improve the quality and intelligibility of the speech signal.This paper is based on multi-channel linear prediction(MCLP)speech dereverberation algorithm,have done the following parts of the research:The traditional weighted prediction error(WPE)algorithm is analyzed.The traditional WPE algorithm is to statistically model the expected Short-Term Fourier Transform coefficients of the original speech signal and find that it conforms to the Time-Varying Gaussian model,and then use the maximum likelihood estimation method to find the desired speech signal.The improved algorithm is to use the Generalized Gaussian Distribution prior probability model which is closer to the real speech signal to estimate the expected speech signal.This paper studies the improved WPE algorithm and finds that when the maximum likelihood estimation method is used to find the unknown parameters,the estimation of the variance of the probability model is not accurate enough.Therefore,Non-negative Matrix Factorization method is introduced,and I-S divergence is used to solve the non-negative matrix factorization optimized variance parameters,thereby improving the dereverberation effect.According to the sparse characteristics of the speech signal,the sparse reconstruction method is used to solve the prediction filter coefficients in the WPE algorithm.The method is to rewrite the expected speech signal into a sparse representation model,and use the Iteratively Reweighted Least Squares(IRLS)algorithm to solve the objective function.This paper studies the solution process of the algorithm,and finds that if the microphone matrix is relatively tight,it will cause the observation matrix to produce singular values and reduce the accuracy of the sparse representation model.After analyzing the advantages and disadvantages of the Ridge regression model and the Lasso regression model,this paper uses the Elastic Network(EN)regression model to constrain the original objective function.Finally,the dereverberation effect of the above two improved algorithms is verified by MATLAB simulation experiments.The Image Source Model algorithm is used to simulate different degrees of room impulse response,and the speech signal is convolved to obtain a reverb voice signal,and the above algorithm is used to dereverberate the reverberation speech signal.By evaluating the dereverberation effect of enhanced speech signals,it can be seen that the algorithm proposed in this paper improves the dereverberation capability of the original algorithm.