请问这个kalman 输出是哪个算法？

Question

请问这个kalman 输出是哪个算法？

zuowanbushiwo opened this issue 2 years ago · comments

zuowanbushiwo commented 2 years ago

是时域 kalman 还是频域的FDAF，还是分块频域PBFDAF？

xuxiaoyu · Answer 1 · Wed Apr 06 2022 17:18:11 GMT+0800 (China Standard Time)

频域分块kalman滤波

zuowanbushiwo · Answer 2 · Wed Apr 06 2022 18:49:16 GMT+0800 (China Standard Time)

是这篇 State-space architecture of the partitioned-block-based acoustic echo controller 文章吗，是不是有用RES？
为什么我仿真的结果不太好，是有什么技巧吗？ process noise covariance 和 measurement noise variance 这两个变量不知道怎么才能比较好的估计出来，在PBFDKF中特别别是 process noise covariance，您有什么好的文章吗？

上面的是你的，下面是我仿真的，不知道是不是代码不对？

function [en,w0,misalign] = PBFDKF(un,dn,M,P,b)
 A = 0.99;
 A2 = A^2;
 L = M/P;
 measurement_smoothing_factor = 0.5; 
 Mk = zeros(2*L,1);
 window_save_first = [ones(L,1); zeros(L,1)];
 window_save_first = repmat(window_save_first,1,P);
 Pk = ones(2*L,P)*10;
 un_blocks = zeros(2*L,1);
 W = zeros(2*L,P);
 loop = floor(length(dn)/L);
 en = zeros(size(dn));
 misalign = zeros(loop,1);
 ComputeEML = 0;
 Xk = zeros(2*L,P);
 w_old = zeros(M,1);
 if nargin == 5 && length(b) == M
      ComputeEML = 1;
 end
 for n = 1:loop
   un_blocks = [un_blocks(L+1:end); un((n-1)*L+1:n*L)]; 
   Uk = fft(un_blocks);
   Xk(:,2:end) = Xk(:,1:end-1);
   Xk(:,1) = Uk;
   y_hat_1 = real(ifft(sum(Xk.*W,2))); 
   y_hat = y_hat_1(L+1:end);
   en((n-1)*L+1:n*L) = dn((n-1)*L+1:n*L)-y_hat;
   
   Ek = fft([zeros(L,1);en((n-1)*L+1:n*L)]);
   Mk = measurement_smoothing_factor*Mk + (1-measurement_smoothing_factor)*abs(Ek).^2;
   Rk = sum(Xk.*Pk.*conj(Xk),2)+2*Mk/P+1e-11;
   mu = (Pk)./Rk;
   K =  mu.*conj(Xk);
   Pk = A2*(1-0.5*K.*Xk).*Pk + (1-A2).*abs(W).^2;
   W = A*(W + fft(window_save_first.*real(ifft(K.*Ek))));
   %%%%%%%residual echo suppression
   % H = 1 - sum(mu.*abs(Xk).^2,2);
   % Ek_Res = H.*Ek;
   % en_t = real(ifft(Ek_Res));
   % en((n-1)*L+1:n*L) = en_t(L+1:end);
   temp = real(ifft(W));
   w0 = temp(1:L,:);
   w0 = w0(:);
   if ComputeEML == 1
       misalign(n) = norm(b-w0)/norm(b); 
   else
       misalign(n) = norm(w_old-w0); 
       w_old = w0;       
   end 
      
 end
 
 if ComputeEML == 1
  figure;
  c = stem([b,w0]);
  legend('Actual','Estimated');
  title('PBFDKF System identification of an FIR filter'); grid on;

  figure;
  hold on; plot((0:length(misalign)-1),10*log10(misalign.^2));
  xlabel('Number of iterations '); 
  ylabel('Misalignment (dB)');
  title('PBFDKF');
  grid on;
  end
end

dragnDriver · Answer 3 · Mon Apr 18 2022 17:57:58 GMT+0800 (China Standard Time)

是这篇 State-space architecture of the partitioned-block-based acoustic echo controller 文章吗，是不是有用RES？为什么我仿真的结果不太好，是有什么技巧吗？ process noise covariance 和 measurement noise variance 这两个变量不知道怎么才能比较好的估计出来，在PBFDKF中特别别是 process noise covariance，您有什么好的文章吗？

上面的是你的，下面是我仿真的，不知道是不是代码不对？

function [en,w0,misalign] = PBFDKF(un,dn,M,P,b)
 A = 0.99;
 A2 = A^2;
 L = M/P;
 measurement_smoothing_factor = 0.5; 
 Mk = zeros(2*L,1);
 window_save_first = [ones(L,1); zeros(L,1)];
 window_save_first = repmat(window_save_first,1,P);
 Pk = ones(2*L,P)*10;
 un_blocks = zeros(2*L,1);
 W = zeros(2*L,P);
 loop = floor(length(dn)/L);
 en = zeros(size(dn));
 misalign = zeros(loop,1);
 ComputeEML = 0;
 Xk = zeros(2*L,P);
 w_old = zeros(M,1);
 if nargin == 5 && length(b) == M
      ComputeEML = 1;
 end
 for n = 1:loop
   un_blocks = [un_blocks(L+1:end); un((n-1)*L+1:n*L)]; 
   Uk = fft(un_blocks);
   Xk(:,2:end) = Xk(:,1:end-1);
   Xk(:,1) = Uk;
   y_hat_1 = real(ifft(sum(Xk.*W,2))); 
   y_hat = y_hat_1(L+1:end);
   en((n-1)*L+1:n*L) = dn((n-1)*L+1:n*L)-y_hat;
   
   Ek = fft([zeros(L,1);en((n-1)*L+1:n*L)]);
   Mk = measurement_smoothing_factor*Mk + (1-measurement_smoothing_factor)*abs(Ek).^2;
   Rk = sum(Xk.*Pk.*conj(Xk),2)+2*Mk/P+1e-11;
   mu = (Pk)./Rk;
   K =  mu.*conj(Xk);
   Pk = A2*(1-0.5*K.*Xk).*Pk + (1-A2).*abs(W).^2;
   W = A*(W + fft(window_save_first.*real(ifft(K.*Ek))));
   %%%%%%%residual echo suppression
   % H = 1 - sum(mu.*abs(Xk).^2,2);
   % Ek_Res = H.*Ek;
   % en_t = real(ifft(Ek_Res));
   % en((n-1)*L+1:n*L) = en_t(L+1:end);
   temp = real(ifft(W));
   w0 = temp(1:L,:);
   w0 = w0(:);
   if ComputeEML == 1
       misalign(n) = norm(b-w0)/norm(b); 
   else
       misalign(n) = norm(w_old-w0); 
       w_old = w0;       
   end 
      
 end
 
 if ComputeEML == 1
  figure;
  c = stem([b,w0]);
  legend('Actual','Estimated');
  title('PBFDKF System identification of an FIR filter'); grid on;

  figure;
  hold on; plot((0:length(misalign)-1),10*log10(misalign.^2));
  xlabel('Number of iterations '); 
  ylabel('Misalignment (dB)');
  title('PBFDKF');
  grid on;
  end
end

不清楚作者是否有做后处理，如果没有做后处理，real_linear_out_kalman.wav这个文件一出来滤波器就收敛了，感觉又有点不对

maoyapeng · Answer 4 · Sun May 15 2022 22:18:51 GMT+0800 (China Standard Time)

同样没做出那个效果

ydaj · Answer 5 · Tue May 24 2022 16:51:28 GMT+0800 (China Standard Time)

想请教您一下关于分块频域Kalman回声消除初始阶段收敛问题您有好的解决办法或者可以提供的思路吗？关于两个噪声估计有没有一些论文推荐呢？谢谢您

zuowanbushiwo · Answer 6 · Fri May 27 2022 09:11:29 GMT+0800 (China Standard Time)

@ydaj 其实我也不知道，我的卡尔曼的效果也不太好。

这篇文章 Robust Uncertainty Control of the Simplified Kalman Filter for Acoustic Echo Cancelation 有介绍一种新的噪声方差估计，不过我实际数据的效果不好，不知道是实现错了。如果你的效果比较好，可以告诉我一下。

maoyapeng · Answer 7 · Tue Jul 19 2022 23:07:48 GMT+0800 (China Standard Time)

路径突变有没有什么解决方法

maoyapeng · Answer 8 · Wed Sep 28 2022 23:30:29 GMT+0800 (China Standard Time)

参考State-space architecture of the partitioned-block-based acoustic echo controller做的效果
滤波器长度4096
纯线性结果：

线性+res效果

低频回声比较严重，有什么方法改善吗

changxu · Answer 9 · Tue Apr 04 2023 16:49:08 GMT+0800 (China Standard Time)

是这篇 State-space architecture of the partitioned-block-based acoustic echo controller 文章吗，是不是有用RES？为什么我仿真的结果不太好，是有什么技巧吗？ process noise covariance 和 measurement noise variance 这两个变量不知道怎么才能比较好的估计出来，在PBFDKF中特别别是 process noise covariance，您有什么好的文章吗？

上面的是你的，下面是我仿真的，不知道是不是代码不对？

function [en,w0,misalign] = PBFDKF(un,dn,M,P,b)
 A = 0.99;
 A2 = A^2;
 L = M/P;
 measurement_smoothing_factor = 0.5; 
 Mk = zeros(2*L,1);
 window_save_first = [ones(L,1); zeros(L,1)];
 window_save_first = repmat(window_save_first,1,P);
 Pk = ones(2*L,P)*10;
 un_blocks = zeros(2*L,1);
 W = zeros(2*L,P);
 loop = floor(length(dn)/L);
 en = zeros(size(dn));
 misalign = zeros(loop,1);
 ComputeEML = 0;
 Xk = zeros(2*L,P);
 w_old = zeros(M,1);
 if nargin == 5 && length(b) == M
      ComputeEML = 1;
 end
 for n = 1:loop
   un_blocks = [un_blocks(L+1:end); un((n-1)*L+1:n*L)]; 
   Uk = fft(un_blocks);
   Xk(:,2:end) = Xk(:,1:end-1);
   Xk(:,1) = Uk;
   y_hat_1 = real(ifft(sum(Xk.*W,2))); 
   y_hat = y_hat_1(L+1:end);
   en((n-1)*L+1:n*L) = dn((n-1)*L+1:n*L)-y_hat;
   
   Ek = fft([zeros(L,1);en((n-1)*L+1:n*L)]);
   Mk = measurement_smoothing_factor*Mk + (1-measurement_smoothing_factor)*abs(Ek).^2;
   Rk = sum(Xk.*Pk.*conj(Xk),2)+2*Mk/P+1e-11;
   mu = (Pk)./Rk;
   K =  mu.*conj(Xk);
   Pk = A2*(1-0.5*K.*Xk).*Pk + (1-A2).*abs(W).^2;
   W = A*(W + fft(window_save_first.*real(ifft(K.*Ek))));
   %%%%%%%residual echo suppression
   % H = 1 - sum(mu.*abs(Xk).^2,2);
   % Ek_Res = H.*Ek;
   % en_t = real(ifft(Ek_Res));
   % en((n-1)*L+1:n*L) = en_t(L+1:end);
   temp = real(ifft(W));
   w0 = temp(1:L,:);
   w0 = w0(:);
   if ComputeEML == 1
       misalign(n) = norm(b-w0)/norm(b); 
   else
       misalign(n) = norm(w_old-w0); 
       w_old = w0;       
   end 
      
 end
 
 if ComputeEML == 1
  figure;
  c = stem([b,w0]);
  legend('Actual','Estimated');
  title('PBFDKF System identification of an FIR filter'); grid on;

  figure;
  hold on; plot((0:length(misalign)-1),10*log10(misalign.^2));
  xlabel('Number of iterations '); 
  ylabel('Misalignment (dB)');
  title('PBFDKF');
  grid on;
  end
end

感觉原文的res压的效果有点弱，换成webrtc的res模块效果会好不少。

zuowanbushiwo · Answer 10 · Thu Apr 06 2023 08:59:11 GMT+0800 (China Standard Time)

你这个是webrtc aec3的效果吗？因为我看你的文件名也是kalman滤波加webrtc的非线性消除，这个和webrtc aec3的组合一样的。