[Speech denoising] Wiener filtering algorithm speech denoising based on matlab prior SNR [with Matlab source code 572]

1. Introduction

This chapter proposes a speech enhancement algorithm based on the Wiener filtering method based on prior signal-to-noise ratio estimation. The initial noise power spectrum is obtained by calculating the statistical average of the silent section, and the initial noise power spectrum and the noisy speech power spectrum are smoothed, and the noise power spectrum is updated. Finally, the situation where the noise increases sharply at a certain frequency point is considered. Correlative verification shows that the algorithm can effectively suppress noise with a small or stable range, but it is not very effective for noise with a wide range of changes in practice.
1. Overview of speech enhancement
1.1 Related concepts of speech enhancement
Embedded in the speech system, the speech signal will inevitably be interfered by the surrounding noise, thereby affecting the quality and intelligibility of the speech.
Speech enhancement: In fact, it is to extract as pure as possible speech from noisy speech, improve speech quality and intelligibility, and improve the performance of speech communication system in noisy environment.
Noise is generated randomly, and it is impossible to completely eliminate it. The goal of speech enhancement is to reduce noise, eliminate background noise, improve speech quality, make the listener receptive, and improve speech intelligibility.
1.2 Related algorithms
for speech enhancement Due to the many sources of noise, the characteristics are different. The application scenarios of speech enhancement processing systems vary greatly.
Therefore, there is no speech enhancement algorithm that can be used in various noisy environments. For different environments, different speech enhancement algorithms are adopted.
According to the processing methods, speech enhancement algorithms can be divided into: enhancement algorithms based on voice periodicity, enhancement algorithms based on all-pole models, enhancement algorithms based on short-time spectrum estimation, enhancement algorithms based on signal subspace, and enhancement algorithms
based on HMM.
Judging from the current development, the method based on short-time spectrum estimation is the most effective method. Specifically, it includes spectral subtraction, Wiener filtering, minimum mean square error short-time spectral amplitude estimation method (MMSE-STSA) and minimum mean square error logarithmic spectral amplitude estimation method (MMSE-LSA). This article mainly discusses the use of Wiener filters to achieve speech enhancement processing.
2 Wiener filter speech enhancement theory based on prior signal-to-noise ratio estimation
The prior signal-to-noise ratio is a very important parameter in the speech enhancement algorithm. The method of calculating the prior signal-to-noise ratio through the "direct decision" estimation proposed by Ephraim and Malah is the most effective and easiest to calculate.
Insert picture description here

Second, the source code

clear all; clc; close all;

[xx, fs] = wavread('C5_3_y.wav');           % 读入数据文件
xx=xx-mean(xx);                         % 消除直流分量
x=xx/max(abs(xx));                      % 幅值归一化
IS=0.25;                                % 设置前导无话段长度
wlen=200;                               % 设置帧长为25ms
inc=80;                                 % 设置帧移为10ms
SNR=5;                                  % 设置信噪比SNR
NIS=fix((IS*fs-wlen)/inc +1);           % 求前导无话段帧数
alpha=0.95;

signal=awgn(x,SNR,'measured','db');               % 叠加噪声
output=Weina_Im(x,wlen,inc,NIS,alpha) ;
output=output/max(abs(output));
len=min(length(output),length(x));
x=x(1:len);
signal=signal(1:len);
output=output(1:len);

snr1=SNR_Calc(x,signal);            % 计算初始信噪比
snr2=SNR_Calc(x,output);            % 计算降噪后的信噪比
snr=snr2-snr1;
fprintf('snr1=%5.4f   snr2=%5.4f   snr=%5.4f\n',snr1,snr2,snr);

% 作图
time=(0:len-1)/fs;                        % 设置时间
subplot 311; plot(time,x,'k'); grid; axis tight;
title('纯语音波形'); ylabel('幅值')
subplot 312; plot(time,signal,'k'); grid; axis tight;
title(['带噪语音 信噪比=' num2str(SNR) 'dB']); ylabel('幅值')
function frameout=enframe(x,win,inc)

nx=length(x(:));            % 取数据长度
nwin=length(win);           % 取窗长
if (nwin == 1)              % 判断窗长是否为1，若为1，即表示没有设窗函数
   len = win;               % 是，帧长=win
else
   len = nwin;              % 否，帧长=窗长
end
if (nargin < 3)             % 如果只有两个参数，设帧inc=帧长
   inc = len;
end
nf = fix((nx-len+inc)/inc); % 计算帧数
frameout=zeros(nf,len);            % 初始化
indf= inc*(0:(nf-1)).';     % 设置每帧在x中的位移量位置
inds = (1:len);             % 每帧数据对应1:len
frameout(:) = x(indf(:,ones(1,len))+inds(ones(nf,1),:));   % 对数据分帧
if (nwin > 1)               % 若参数中包括窗函数，把每帧乘以窗函数
    w = win(:)';            % 把win转成行数据
    function frameout=filpframe(x,win,inc)

[nf,len]=size(x);
nx=(nf-1) *inc+len;                 %原信号长度
frameout=zeros(nx,1);
nwin=length(win);                   % 取窗长
if (nwin ~= 1)                           % 判断窗长是否为1，若为1，即表示没有设窗函数
    winx=repmat(win',nf,1);
    x=x./winx;                          % 除去加窗的影响
    x(find(isinf(x)))=0;                %去除除0得到的Inf
end

 
for i=1:nf
    start=(i-1)*inc+1;    
    xn=x(i,:)';
    sig(start:start+len-1)=sig(start:start+len-1)+xn;
end

Three, running results

Insert picture description here

Four, remarks