vc1.m

这是一个用于语音信号处理的工具箱

   out=lrest(ngm1_ref, ngm2_ref, 1, SHOW)
   b=out(1);
   a=out(2);
   ngm3_ref=abs(a*ngm1_ref+b);  % new ngm contour
   
   x_ref=road(1,:);
   tM2=length(x_ref);
   tmp_gm=zeros(tM2,4);
   
   % ---interpolate ngm3---
   s_fram=1;
   for k=1:length(vcidx21);
      e_fram=s_fram+vcidx22(k)-vcidx21(k);
      
      % smoothing the starting frame
      kf=s_fram;
      gm_x=[1 m_len/2 m_len+m_len];
      %When using segmentation we don't want to start the gain contour of the segment at 0
      if ~exist('seg','var')   
         ngm3_ref(kf)=0.13*ngm3_ref(kf+1);
         gm_y=[0 ngm3_ref(kf) ngm3_ref(kf+1)];
      else
         gm_y=[ngm3_ref(kf) ngm3_ref(kf) ngm3_ref(kf+1)];
      end
      gm_z=m_len/4*[0 1 2 3]+m_len/8;
      tmp=interp1(gm_x,gm_y,gm_z,'spline');
      tmp_gm(kf,1:4)=tmp(:)';
      
      for kf=s_fram+1:e_fram-1
         gm_x=[-1*m_len/2 m_len/2 m_len+m_len];
         gm_y=[ngm3_ref(kf-1) ngm3_ref(kf) ngm3_ref(kf+1)];
         gm_z=m_len/4*[0 1 2 3]+m_len/8;
         tmp=interp1(gm_x,gm_y,gm_z);
         tmp_gm(kf,1:4)=tmp(:)';
      end
      
      % smoothing the ending frame
      kf=e_fram;
      gm_x=[-1*m_len/2 m_len/2 m_len];
      %When using segmentation we don't want to end the gain contour of the segment at 0
      if ~exist('seg','var')
         ngm3_ref(kf)=0.13*ngm3_ref(kf-1);
         gm_y=[ngm3_ref(kf-1) ngm3_ref(kf) 0];
      else 
         gm_y=[ngm3_ref(kf-1) ngm3_ref(kf) ngm3_ref(kf)];
      end
      gm_z=m_len/4*[0 1 2 3]+m_len/8;
      tmp=interp1(gm_x,gm_y,gm_z,'spline');
      tmp_gm(kf,1:4)=tmp(:)';
      
      s_fram=e_fram+1;
   end  %% for k=1:length(vcidx21)
   
   ngm3( x_ref,1:4)=tmp_gm;
   
   disp('The linear model is used to convert the gain contour.');
   disp( [a b]);
   disp('Normalized Standard Deviation=');
   disp(std(ngm3_ref-ngm2_ref)/mean(ngm2_ref));
   
elseif ngm_method==3
   
   ngm3_ref=ngm2_ref;
   
elseif ngm_method==4
   
   ngm3_ref=ngm1_ref;
   x_ref=road(1,:);
   tM2=length(x_ref);
   tmp_gm=zeros(tM2,4);
   
   % ---interpolate ngm3---
   s_fram=1;
   for k=1:length(vcidx21);
      e_fram=s_fram+vcidx22(k)-vcidx21(k);
      
      % smoothing the starting frame
      kf=s_fram;
      ngm1_ref(kf)=0.13*ngm1_ref(kf+1);
      gm_x=[1 m_len/2 m_len+m_len];
      gm_y=[0 ngm1_ref(kf) ngm1_ref(kf+1)];
      gm_z=m_len/4*[0 1 2 3]+m_len/8;
      tmp=interp1(gm_x,gm_y,gm_z,'spline');
      tmp_gm(kf,1:4)=tmp(:)';
      
      for kf=s_fram+1:e_fram-1
         gm_x=[-1*m_len/2 m_len/2 m_len+m_len];
         gm_y=[ngm1_ref(kf-1) ngm1_ref(kf) ngm1_ref(kf+1)];
         gm_z=m_len/4*[0 1 2 3]+m_len/8;
         tmp=interp1(gm_x,gm_y,gm_z);
         tmp_gm(kf,1:4)=tmp(:)';
      end
      
      % smoothing the ending frame
      kf=e_fram;
      ngm1_ref(kf)=0.13*ngm1_ref(kf-1);
      gm_x=[-1*m_len/2 m_len/2 m_len];
      gm_y=[ngm1_ref(kf-1) ngm1_ref(kf) 0];
      gm_z=m_len/4*[0 1 2 3]+m_len/8;
      tmp=interp1(gm_x,gm_y,gm_z,'spline');
      tmp_gm(kf,1:4)=tmp(:)';
      
      s_fram=e_fram+1;
   end  %% for k=1:length(vcidx21)
   
   ngm3( x_ref,1:4)=tmp_gm;
   
end  %% if ngm_method==1

clear ngm1_ref ngm2_ref tmp tmp_gm;


%-------------------------------------------------
%************ convert gm3 to match gm2  **********

% construct new gain contour 

gm_x=1;
gm_y=0;
[nframe,dum]=size(cofa2);
for kf=1:nframe
   spoint=m_len*(kf-1)+Order+1;
   tmp_x=m_len/4*[0 1 2 3]+m_len/8+spoint;
   gm_x=[gm_x tmp_x];
   gm_y=[gm_y ngm3(kf,1:4)];
end
gm_x=[gm_x spoint+m_len-1];
gm_y=[gm_y 0];

gm3=interp1( gm_x,gm_y,gci3 );
gm3=gm3(:)';
ave=min(gm3);
gm3=filter(0.4,[1 -0.4 -0.2],gm3);
gm3=gm3+(ave-min(gm3));
gm3=abs(gm3);

% take care of the transition. In the case of segmantation it does not
if length(gm3)==length(gm1)
   gm3(1)=gm1(1);
   gm3(length(gm3))=gm1(length(gm1));
%else
%   [y,cor23]=min(abs(gci2-gci3(1)));
%   gm3(1)=gm2(cor23);
%   [y,cor23]=min(abs(gci2-gci3(length(gci3))));
%   gm3(length(gm3))=gm2(cor23);
end
if SHOW==1
   plot(gci1,gm1,'y',gci3,gm3,'r',gci2,gm2,'g');
   title('gain conversion (source: yellow ; converted: red ; target: green)');
   pause(1);
end

clear gm_x gm_y gm_z tmp_x cor23 y;

%************ convert gpcf1 to match gpcf2  **********

gpcf3=zeros(M2,7);
gpcf_method=get(v_pp_glo,'Value');

if sr==1
   ncof=7;  % number of coefficients
elseif sr==2
   ncof=5;
end

if gpcf_method==1
   
   for kk=1:ncof
      out=lrest(gpcf1_ref(:,kk), gpcf2_ref(:,kk), 0, SHOW);
      b=out(1);
      gpcf1_ref(:,kk)=gpcf1_ref(:,kk)+b;
      err(kk)=out(3);
   end
   
   gpcf3( x_ref,:)=gpcf1_ref;
   disp('The bias model is used to convert glottal source.');
   disp(b)
   disp('Normalized Standard Deviation=')
   disp(err./mean(gpcf2_ref(:,1:ncof)));
   
elseif gpcf_method==2
   
   disp('The linear model is used to convert the glottal source.');
   
   for kk=1:ncof
      out=lrest(gpcf1_ref(:,kk), gpcf2_ref(:,kk), 1, SHOW);
      b=out(1);
      a=out(2);
      err(kk)=out(3);
      gpcf1_ref(:,kk)=a*gpcf1_ref(:,kk)+b;
      disp( [a b]);
   end
   disp('Normalized Standard Deviation=');
   disp(err./mean(gpcf2_ref(:,1:ncof)));
   
   gpcf3( x_ref,:)=gpcf1_ref;
   
elseif gpcf_method==3
   
   gpcf3=gpcf2;
   
elseif gpcf_method==4
   
   gpcf3( x_ref,:)=gpcf1_ref;
   
end

clear gpcf1_ref gpcf2_ref tmp ncof err;

%************ convert FF1 to match FF2  **********

ff_method=get(v_pp_ff,'Value');

if vt==1 & ff_method==4        
   cofa3(road(1,:),:)=cofa1(road(2,:),:);
   
end  %% if vt==1

if vt==2
   
   x_ref=road(1,:);
   FF3=FF2;
   
   if ff_method==1
      nff=basic(1);
      for kk=1:nff
         out=lrest(FF1_ref(:,kk), FF2_ref(:,kk), 0, SHOW);
         b=out(1);
         err(kk)=out(3);
         FF3_ref(:,kk)=FF1_ref(:,kk)+b;
      end
      
      FF3( x_ref,:)=FF3_ref;
      disp('The bias model is used to convert the formant track.');
      disp(b);
      disp('Normalized Standard Deviation=');
      disp(std(FF3_ref-FF2_ref)./mean(FF2_ref));
      
   elseif ff_method==2
      
      disp('The linear model is used to convert the formant track.');
      nff=basic(1);
      for kk=1:nff
         out=lrest(FF1_ref(:,kk), FF2_ref(:,kk), 1, SHOW);
         b=out(1);
         a=out(2);
         err(kk)=out(3);
         FF3_ref(:,kk)=a*FF1_ref(:,kk)+b;
         disp( [a b]);
      end
      disp('Normalized Standard Deviation=');
      disp(err./mean(FF2_ref));
      
      FF3( x_ref,:)=FF3_ref;
      
   elseif ff_method==3
      
      FF3=FF2;
      FB3=FB2;
      FF3_ref=FF2_ref;
      
   elseif ff_method==4
      
      FF3_ref=FF1_ref;
      FB3=FB2;
      FF3( x_ref,:)=FF3_ref;
      FB3( x_ref,:)=FB1(road(2,:),:);       
      cofa3(road(1,:),:)=cofa1(road(2,:),:);
      
   end
   
   clear s_fram e_fram spoint epoint err;
   
   %---------------------------------------%
   % determine new formant bandwidth FB3   %
   %---------------------------------------%
   
   % change the formant bandwidth to compensate for the pole interaction problem
   if ff_method<3
      FB3=FB2;
      for k=1:tM2
         kf1=road(2,k);
         kf2=road(1,k);
         %[npoly,opoly,FB3(kf2,:)]=rt2lp(FF3(kf2,:),FF1(kf1,:),FB1(kf1,:) );
         [npoly,opoly,FB3(kf2,:)]=rt2lp(FF3(kf2,:),FF2(kf2,:),FB2(kf2,:) );
      end
      
   end %%if ff_method<3
   
   if SHOW==1         
      subplot(311); plot(FF1_ref);title('source formant track');
      subplot(312); plot(FF3_ref);title('converted formant track');
      subplot(313); plot(FF2_ref);title('target formant track');
      pause(1);
      subplot(111);
   end
end %% if vt==2;

disp('Voice Conversion is completed. You may run the synthesizer.');
clear tmp apply FF1_ref FF2_ref FF3_ref  x_ref gci_s gci_e kf1 kf2 out err;
%clear vcidx11 vcidx12 vcidx21 vcidx22;
%clear ngm1 gm1 gci1 gpcf1 FF1 FB1 road road1 cofa1 vctyp1 nidx1