www.pudn.com > formant_track.zip > lpcsgram.m

function [tfr,fmt,pt] = lpcsgram(sig,Nt,Nf,fs) 
% LPCSGRAM Determine the spectrogram from an lpc analysis of the speech signal 
%   sig : Input signal 
%   Nt  : Number of time points 
%   Nf  : Number of frequency pts 
%   fs  : Sampling frequency 
%   The outputs are: 
%   tfr : Time frequency image [NfxNt] 
%   fmt : Fmt tracks [3xNt] 
%   pt  : Pitch track 
 
if nargin<4, 
    fs = 11025; 
end; 
 
% Number of lpc coefficients 
Nlpc = round(fs/1000)+2; 
 
% Window width 
Nwin = floor(length(sig)/Nt); 
 
tfr = zeros(Nf,Nt); 
for i=1:Nt, 
    % windowed signal 
    lpcsig = sig((Nwin*(i-1)+1):min([(Nwin*i) length(sig)])); 
     
    if ~isempty(lpcsig), 
        % perform lpc analysis 
        [A,G] = lpc(lpcsig,Nlpc); 
        e = lpcsig-real(filter([0 -A(2:end)],1,lpcsig)); 
        hf = freqz(sqrt(G),A,Nf); 
         
        % Determine the autocorrelation function 
        acf = xcorr(e,Nwin,'coeff'); 
 
        % Divide the symmetric function in half, excluding peak 
        % at the midpoint 
        acf = acf((fix(end/2)+2):end); 
        idx = find(acf>0.25); 
         
        % Determine pitch for the window 
        if isempty(idx) | (length(idx)==1 & idx<(fs/300)) | Nwin<100, 
            pt(i) = NaN; 
        else 
            if idx(1)<(fs/300), 
                pt(i) = fs/idx(2); 
            else, 
                pt(i) = fs/idx(1); 
            end; 
        end 
         
        % Determine the formants 
        %if G>1.5e-4, 
        fmt(:,i) = [frmnts1(A(:),fs)/(fs/2)]; 
        %else, 
        %fmt(:,i) = [NaN;NaN;NaN]; 
        %end; 
         
        % Create a slice of the time-frequency image 
        tfr(:,i) = abs(hf(:)); %/norm(abs(hf(:))); 
    end; 
end;