1. Software version

matlab2021a

2. Theoretical knowledge of this algorithm

        The most common technique for dealing with broadband noise is spectral subtraction , That is, the noise spectrum estimation is subtracted from the noisy speech estimation , And get the spectrum of pure speech . Because the human ear is insensitive to the phase of speech spectral components , Therefore, this method is mainly aimed at short-time amplitude spectrum . The speech is assumed to be a stationary signal , Noise and speech are additive signals and are not related to each other . At this time, the noisy speech signal can be expressed as .

3. Part of the core code

function enhancedsignal=wiener(noisyspeech,samplefrequency)
x=noisyspeech;
fs=samplefrequency;
nx=length(x);
enhanced_x=zeros(1,nx);
                                 % Framing and windowing 
FrameLen=fix(0.025*fs);          % take 25 Milliseconds is a frame 
overlap=FrameLen/2;
inc=FrameLen-overlap;            % Frame shift 
x_frame=enframe(x,FrameLen,inc); % Framing 
nf=size(x_frame,1);              %  frames 
win=hamming(FrameLen)';
x_window=[];
for k=1:nf
    x_row=x_frame(k,:).*win;     %  Add windows 
    x_window=[x_window;x_row];   
end
% For noisy speech DFT
y=fft(x_window');
ymag = abs(y);			
yphase = angle(y);	                         
NNoise=23;                      % Take the noise section （ The initial segment of speech ） frames 
MN=mean(ymag(:,1:NNoise)')';
PN=mean(ymag(:,1:NNoise)'.^2)'; % Mean value of initial noise power spectrum 

NoiseCounter=0;% Length of continuous noise section 
SmoothFactor=9;% Noise smoothing factor 
Alpha=0.95;    % Speech smoothing factor 
SNRPre=ones(size(MN));

% Wiener filtering 
for k=1:nf
     if k<=NNoise 
        SpeechFlag=0;
        NoiseCounter=NNoise;
    else 
           NoiseMargin=3;
           HangOver=8;
           SpectralDist= 20*(log10(ymag(:,k))-log10(MN));
           SpectralDist(find(SpectralDist<0))=0;
           Dist=mean(SpectralDist); 
           if (Dist < NoiseMargin) 
            NoiseFlag=1; 
            NoiseCounter=NoiseCounter+1;
           else
            NoiseFlag=0;
            NoiseCounter=0;
           end 
           if (NoiseCounter > HangOver) 
           SpeechFlag=0;    
           else 
           SpeechFlag=1; 
           end 
     end
    
    if SpeechFlag==0 
        MN=(SmoothFactor*MN+ymag(:,k))/(SmoothFactor+1);      % Update the noise mean 
        PN=(SmoothFactor*PN+(ymag(:,k).^2))/(1+SmoothFactor); % Update noise power 
    end
    
%------ wave filtering 
SNRNew=(ymag(:,k).^2)./PN-1;
SNRPost=Alpha*SNRPre+(1-Alpha).*max(SNRNew,0);
Gain=SNRPost./(SNRPost+1);
smag=Gain.*ymag(:,k);
SNRPre=smag.^2./PN;
spectrum= smag.*exp(j*yphase(:,k));
enhanced_x((inc*(k-1)+1):(inc*(k-1)+FrameLen))=enhanced_x((inc*(k-1)+1):(inc*(k-1)+FrameLen))+real(ifft(spectrum,FrameLen))';
end
enhancedsignal=enhanced_x;

4. Simulation conclusion

5. reference

[1] Yi Kechu . Voice signal processing [M]. Beijing : National Defense Industry Press , 2000.

[2] LiZhao, KOBAYASHI, NIIMI Y Tone. Recongnition of Chinese continuous speech using continuous HMMs[J]. Journal of the Acoustical Society of Japan, 53(12), 933-940, 1997.

A03-03

6. How to obtain the complete source code

The way 1： Wechat or QQ Contact bloggers
The way 2： subscribe , Get the tutorial case code and any of this blog for free 2 Complete source code