lpana3.m

来自「这是一个用于语音信号处理的工具箱」· M 代码 · 共 264 行

% Function: preform Linear Prediction Speech Analysis.

function [gm,gpcof,ncidx,ncgm]=lpana3(signal,basic,gci,voicetype,rsd,cofa);

%retrieve the basic specification
F_len=basic(5);
O_lap=basic(6); 
Order=basic(4);
M_len=F_len-O_lap;
[nframe,dum]=size(cofa);


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%                           %%
%%%%%%%  Find glottal codeword    %%
%%%%%%%                           %%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

 % This step will create the following variables :
 %
 %	gpcof = glottal codeword for each pitch period.

 %-----------------------------%
 % construct the ploc1 matrix  %
 %-----------------------------%

 ploc1=zeros(nframe,10);
 k_back=0;
 for i=1:length(gci)
     k=fix( (gci(i)-Order)/M_len +1);
     aa=gci(i)-Order-(k-1)*M_len;
     if aa==0 
        k=k-1;
        aa=M_len;
     end
     if (k-k_back)==0
        idx=idx+1;
     else
        idx=1;
     end  

     ploc1(k,idx)=aa;
     k_back=k;
 end

gpcof=zeros(nframe,7);

% Smooth integrated residue.
ir1=integ(rsd);  
ir1=filtfilt([1 -1],[1 -.99],ir1);
%ir=filtfilt([1 -1],conv([1 -.9],[1 -.7]),ir1); 
dgf=ir1;  % estimated differentiated glottal flow
%gf=integ(dgf); % estimated glottal flow

Smodel=basic(3);
          %--- Select Glottal Source model
          %---- Smodel==1 --> 6 order polynomial model (sounds better)
          %---- Smodel==2 --> LF model 

for kf=2:nframe-1
	if ploc1(kf,1)
	    startp=M_len*(kf-1)+Order;
	    self=ploc1(kf,:);
	    self=self(self~=0);
	    if ploc1(kf+1,1)
		after=ploc1(kf+1,1)+M_len;
	    elseif voicetype(kf+1) & kf <= nframe-2 
	      	if ploc1(kf+2,1)
		    after=ploc1(kf+2,1)+2*M_len;
		else
		    after=[];
	        end;
	    else
		after=[];
	    end;
	    period=[self after]+startp;
	    lenp=length(period);
	    if lenp>1
		gcof=[];
		for k=1:lenp-1
			ii=dgf(period(k):period(k+1));

                        % Polynomial model when Smodel==1 ***
                        if Smodel==1
			   gcof(k,:)=polym1(ii);
                           % polym1 <--> gen_dgf1

                        % LF model when Smodel==2 ***
                        elseif Smodel==2
                           ii=integ(ii); % glottal flow
                           [Rd,Ee]=lfmodel(ii);
                           Ra=(-1+4.8*Rd)/100;
                           Rk=(22.4+11.8*Rd)/100;
                           Rg=Rk/4/( 0.11*Rd/(0.5+1.2*Rk)-Ra);
                           T0=length(ii)-1;
                           Ta=T0*Ra;
                           Tp=T0/2/Rg;
                           Te=floor(Tp*Rk+Tp);
                           Tc=T0;
                           gcof(k,1:7)=[Tp/T0 Te/T0 Ta/T0 Tc/T0 Ee 0 0];
                        end
		end;
		if lenp<=2
			avegcof=gcof;
		else
			avegcof=mean(gcof);
		end;
		gpcof(kf,:)=avegcof;

             end %% if lenp>1
	end %% if ploc1(kf,1)
end

%disp('Glottal codeword searching is completed!'); toc;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%                                     %%
%%%%%%%  energy encoding for voiced speech  %%
%%%%%%%                                     %%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% This step will create the following variables :
%
%	gm = signal power for each pitch period.


% encoding the energy for pitch period
numgci=length(gci);
gm=zeros(1,numgci);
if numgci>1
   lenss=gci(2)-gci(1);
end
for nk=1:numgci

       startp=gci(nk)+1;  % starting point of the pitch period
       kf=fix( (startp-Order)/M_len+1 ); % frame number

       if nk==numgci
          endp=startp+lenss-1;
       else
          endp=gci(nk+1);  % ending point of the pitch period
          olenx=lenss;
          lenss=gci(nk+1)-gci(nk);
       end

       if lenss>F_len
          if voicetype(kf)==1 & voicetype(kf+1)==0
             lenss=olenx;
             endp=startp+lenss-1;
          end
       end

       sp=signal(startp:endp);
       gm(nk)=(sp*sp')/lenss;
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%                                        %%
%%%%%%%  energy encoding for the whole speech  %%
%%%%%%%                                        %%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% This step will create the following variables :
%
%    ncgm = signal power for Xlen period; size(ncgm)=[nframe,4]

ncgm=zeros(nframe,4);
Xlen=M_len/4;
for kf=1:nframe
    for kk=1:4
        startp=(kf-1)*M_len+Order+(kk-1)*Xlen+1;
        ss1=signal(startp:startp+Xlen-1);
        ncgm(kf,kk)=(ss1*ss1')/Xlen;
    end
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%                            %%
%%%%%%%  Find unvoiced codeword    %%
%%%%%%%                            %%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

 % This step will create the following variables :
 %
 %	ncidx = glottal codeword.

% Function : preprocess before performing codeword searching.  This includes
%		1. removal of filter memory contribution,
%		2. spectral weighting.

%disp('');
%disp('It now takes a while to perform codeword searching for unvoiced speech');
%disp('Please wait! Please wait! Please wai-------------------------------t.');
%disp('');

load nc;  % Load the stochastic codebook
Xlen=M_len/4;
Ziy1=zeros(1,Order);  % Reset filter memory at the beginning
for kf=1:nframe
	if kf==1 | kf==nframe
		proc='nc';
	elseif sum(voicetype(kf-1:kf))<2 | sum(voicetype(kf:kf+1))<2
		proc='nc';
	else
		proc='mp';
	end;
	if proc=='nc'  % Proceed if it is an unvoiced segment
	    ocofa=cofa(kf,:);
            ncofa=ocofa.*((0.8).^(0:length(ocofa)-1));
	    if kf>1
	    	if voicetype(kf-1)
			Ziy1=signal((kf-1)*M_len+Order:-1:(kf-1)*M_len+1);
			% Note : filter memory is regarded as the past samples
			% for the Direct-I implementation
		end;
	    end;
	    if kf==1 
		so=signal((kf-1)*M_len+Order+1:kf*M_len+Order);
	    else
		so=signal((kf-1)*M_len+Order+1:kf*M_len+Order)-...
		    filt(1,ocofa,zeros(1,M_len),Ziy1);
		% Remove filter memory carried over from the previous frame
	    end
	    ss=filter(ocofa,ncofa,so);  % spectral weighting 
	    % PURPOSE: De-emphasize errors at the formant regions
	    noise1=[];
	    Ziy=Ziy1;  % filter memory

	    for kk=1:4
		cmp1=ss(1+(kk-1)*Xlen:Xlen*kk);
		sub1=filt(1,ncofa,zeros(1,Xlen),Ziy);
		cmp=cmp1-sub1;  % eliminate memory contribution
		ncidx(kf,kk)=fdncb0(cmp,ncofa,nc0,Rnc0);

                % track the filter memory
                eng1=ncgm(kf,kk);
		noise=nc0(ncidx(kf,kk),:);
                if Xlen~=50
                   noise=interpft(noise,Xlen);
                end
                ss2=filt(1,ocofa,noise,Ziy);
                eng2=sum(ss2.^2);
                amp=sqrt(eng1/eng2);

		% Concatenate the excitation in every subframe as a whole
		noise1=[noise1 amp*noise];
	    	nsyn=amp*ss2;
		Ziy=nsyn(Xlen:-1:Xlen-Order+1);
	    end
	    ss1=filt(1,ocofa,noise1,Ziy1);
	    Ziy1=ss1(M_len:-1:M_len-Order+1);  % Track the filter memory
	end	
end

clear Rnc0 nc0 ss1 ss2 noise noise1;

%disp('Analysis is OK!');toc;

%disp('Stochastic codeword searching is completed!'); toc;
%disp('');
%disp('Analysis is OK! You may perform speech synthesis now!');
%disp('');
%disp('');