An Improved Parameter Adaptive Wiener Filtering Speech Enhancement Algorithm

In recent years,with the development of the information technology,the hearing aid technology has matured,using smart phones for voice communication has become a necessity in people's lives,and the human-computer interaction also gradually has entered people's field of vision.Speech has become an indispensable part of daily life,therefore,to deal with voice becomes imminent.In real life,in the speech processing,noise is everywhere,and the complex noise becomes the biggest obstacle in the process of voice.In the past few decades,how to separate the clean speech from the speech signals with noise had been the focus in the study of individual researchers.And the speech enhancement is still an active field now.The traditional speech enhancement algorithm includes spectrum subtraction,wiener filtering algorithm,subspace algorithm and minimum mean square error,etc.While carrying on the speech enhancement,all of these traditional algorithms don't consider that different noises will bring different effect on speech signals.In fact,in the process of speech enhancement,there are two important stages,which are speech estimation and noise estimation.In the noise estimation stage,different noises have different spectrum characteristics.Based on this theory,this paper puts forward a parameter adaptive speech enhancement algorithm according to different noises.Firstly,we introduced relevant theory of the speech and noise,including the main characteristics of the speech and noise,the signal model of the noisy speech and the noise assessment algorithm.We summarized and compared the existing speech enhancement algorithms in recent years,conducted the simulation experiments for the classical speech enhancement,and got the common problems of these algorithms——the same kind of speech enhancement algorithms can't work for all types of noise.Secondly,for making the speech enhancement algorithms conducted different treatment for different noises,we need to do the accurate noise classification.In order to accurately classify the noise,after summarizing a series of classical classification algorithms,we stated the existing problem of these classification algorithms on large data processing,And then we put forward the classification algorithm used in this paper——the noise classification algorithms based on deep belief network.We deep introduced the process of this classification algorithm,conducted the simulation experiment for the classification accuracy,and compared the accuracy with the neural network algorithm.The results showed that the classification algorithm proposed by this paper has higher classification accuracy.Thirdly,because the wiener filtering speech enhancement algorithm based on priori-SNR could greatly improve the quality of the speech and contain lessmusic noise,in this paper,we derived and analyzed it in detail,did the corresponding simulation experiment,got the algorithm can not improve the speech quality of all types of noise interference and proposed the algorithms of this paper——the parameter adaptive wiener filtering speech enhancement algorithm based on deep belief network.This paper selected the voice activity detection(VAD)method for estimating the noise power spectrum,and combined the noise power spectrum with the wiener filtering algorithm to get the best parameter combination for the different noises.Then we applied this combination to the noise types which we got from deep belief network,amended the overestimated part for the priori-SNR estimation,and obtained the speech enhancement algorithm which we proposed.Finally,we conducted simulation experiments for the proposed algorithm,and compared it with the classical wiener filtering algorithm based on priori-SNR.The evaluation algorithm included quality assessment and intelligibility assessment,and we choose perceptual evaluation of speech quality(PESQ)for quality assessment and normalized covariance metric(NCM)for intelligibility assessment.At the end of the experiment,we conducted simulation experiment for the unknown noise.In order to prove that the proposed algorithm had universal application,we were using the proposed classification algorithm to match the unknown noise to the known noise and depending on the type of match to choose the optimal parameter combination for speech enhancement.Results showed that the proposed algorithm could improve the speech qualitywithout reducing the low-SNR speech intelligibility,and it had universal application for the unknown noise.