【语音增强】谱减法、最小均方和维纳滤波语音增强matlab源码

一、简介

在语音去噪中最常用的方法是谱减法，谱减法是一种发展较早且应用较为成熟的语音去噪算法，该算法利用加性噪声与语音不相关的特点，在假设噪声是统计平稳的前提下，用无语音间隙测算到的噪声频谱估计值取代有语音期间噪声的频谱，与含噪语音频谱相减，从而获得语音频谱的估计值。谱减法具有算法简单、运算量小的特点，便于实现快速处理，往往能够获得较高的输出信噪比，所以被广泛采用。该算法经典形式的不足之处是处理后会产生具有一定节奏性起伏、听上去类似音乐的“音乐噪声”。

转换到频域后，这些峰值听起来就像帧与帧之间频率随机变化的多频音，这种情况在清音段尤其明显，这种由于半波整流引起的“噪声”被称为“音乐噪声”。从根本上，通常导致音乐噪声的原因主要有：
（1）对谱减算法中的负数部分进行了非线性处理
（2）对噪声谱的估计不准
（3）抑制函数（增益函数）具有较大的可变性
 1 原理

 

 二、源代码

```c
global hr1 hr2 hr3 hr4 s y fs
clf reset
set(gcf,'menubar','none')
set(gcf,'unit','nor malized','position',[0.1,0.1,0.85,0.85]);
set(gcf,'defaultuicontrolunits','normal')
set(gcf,'defaultuicontrolfontsize',12);
uicontrol('style','frame','position',[0.64,0.6,0.3,0.3]);
uicontrol('style','text','string','选择按钮框','position',[0.65,0.91,0.12,0.03],'horizontal','left');
hr1=uicontrol(gcf,'style','popupmenu','string','谱减法|维纳滤波法|最小均方误差估计法','position',[0.65,0.85,0.25,0.03]);
hr2=uicontrol(gcf,'style','toggle','string','开始/关闭','position',[0.72,0.65,0.15,0.05]);
%uicontrol(gcf,'style',')
%htitle1=title('原是语音波形');
uicontrol('style','text','string','原始语音波形','position',[0.25,0.93,0.12,0.03],'horizontal','center');
h_axes1=axes('position',[0.05,0.54,0.52,0.38]);
set(h_axes1,'ylim',[-1,1]);
%t=0:pi/50:2*pi;
%y=sin(t);
%plot(t,y);
[y,fs,bit]=wavread('C:\Users\lenovo\Desktop\89787488speech_enhancement_GUI\speech enhancement\5.wav');
L1=length(y);
t1=1:L1;
plot(t1,y);

uicontrol('style','text','string','增强后语音波形','position',[0.25,0.45,0.12,0.03],'horizontal','center');
h_axes2=axes('position',[0.05,0.05,0.52,0.38]);
set(h_axes2,'ylim',[-1,1]);
set(hr1,'callback','speech_enhancement');
set(hr2,'callback','speech_enhancement');
function [ss,po]=specsubm(s,fs,p)
%SPECSUBM performs speech enhancement using spectral subtraction [SS,PO]=(S,FS,P)
%
% implementation of spectral subtraction algorithm by R Martin (rather slow)
% algorithm parameters: t* in seconds, f* in Hz, k* dimensionless
% 1: tg = smoothing time constant for signal power estimate (0.04): high=reverberant, low=musical
% 2: ta = smoothing time constant for signal power estimate
%        used in noise estimation (0.1)
% 3: tw = fft window length (will be rounded up to 2^nw samples)
% 4: tm = length of minimum filter (1.5): high=slow response to noise increase, low=distortion
% 5: to = time constant for oversubtraction factor (0.08)
% 6: fo = oversubtraction corner frequency (800): high=distortion, low=musical
% 7: km = number of minimisation buffers to use (4): high=waste memory, low=noise modulation
% 8: ks = oversampling constant (4)
% 9: kn = noise estimate compensation (1.5)
% 10:kf = subtraction floor (0.02): high=noisy, low=musical
% 11:ko = oversubtraction scale factor (4): high=distortion, low=musical
if nargin<3 po=[0.04 0.1 0.032 1.5 0.08 400 4 4 1.5 0.02 4].'; else po=p; end
ns=length(s);
ts=1/fs;
ss=zeros(ns,1);

ni=pow2(nextpow2(fs*po(3)/po(8)));
ti=ni/fs;
nw=ni*po(8);
nf=1+floor((ns-nw)/ni);
nm=ceil(fs*po(4)/(ni*po(7)));

win=0.5*hamming(nw+1)/1.08;win(end)=[];
zg=exp(-ti/po(1));
za=exp(-ti/po(2));
zo=exp(-ti/po(5));

px=zeros(1+nw/2,1);
pxn=px;
os=px;
mb=ones(1+nw/2,po(7))*nw/2;
im=0;
osf=po(11)*(1+(0:nw/2).'*fs/(nw*po(6))).^(-1);

imidx=[13 21]';
x2im=zeros(length(imidx),nf);
osim=x2im;
pnim=x2im;
pxnim=x2im;
qim=x2im;

for is=1:nf
   idx=(1:nw)+(is-1)*ni;
   x=rfft(s(idx).*win);
   x2=x.*conj(x);
   
   pxn=za*pxn+(1-za)*x2;
   im=rem(im+1,nm);
   if im
      mb(:,1)=min(mb(:,1),pxn);
   else
      mb=[pxn,mb(:,1:po(7)-1)];
   end
   pn=po(9)*min(mb,[],2);
   %os= oversubtraction factor
   os=zo*os+(1-zo)*(1+osf.*pn./(pn+pxn));
   
   px=zg*px+(1-zg)*x2;
   q=max(po(10)*sqrt(pn./x2),1-sqrt(os.*pn./px)); 
   ss(idx)=ss(idx)+irfft(x.*q);
   
end
```

三、运行结果