gusucode.com > 《MATLAB图像与视频处理实用案例详解》代码 > 《MATLAB图像与视频处理实用案例详解》代码/第 19 章 基于语音识别的信号灯图像模拟控制技术/voicebox/fxpefac.m
function [fx,tx,pv,fv]=fxpefac(s,fs,tinc,m,pp)
%FXPEFAC PEFAC pitch tracker [FX,TT,PV,FV]=(S,FS,TINC,M,PP)
%
% Input: s(ns) Speech signal
% fs Sample frequency (Hz)
% tinc Time increment between frames (s) [0.01]
% or [start increment end]
% m mode
% 'g' plot graph showing waveform and pitch
% 'G' plot spectrogram with superimposed pitch
% 'x' use external files for algorithm parameter
% initialization: fxpefac_g and fxpefac_w
% pp structure containing algorithm parameters
%
% Outputs: fx(nframe) Estimated pitch (Hz)
% tx(nframe) Time at the centre of each frame (seconds).
% pv(nframe) Probability of the frame of being voiced
% fv structure containing feature vectors
% fv.vuvfea(nframe,2) = voiced/unvoiced GMM features
% References
% [1] S.Gonzalez and M. Brookes,
% A pitch estimation filter robust to high levels of noise (PEFAC), Proc EUSIPCO,Aug 2011.
% Bugs/Suggestions
% (1) do long files in chunks
% (2) option of n-best DP
% Copyright (C) Sira Gonzalez and Mike Brookes 2011
% Version: $Id: fxpefac.m,v 1.2 2011/07/27 07:22:07 dmb Exp $
%
% VOICEBOX is a MATLAB toolbox for speech processing.
% Home page: http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/voicebox.html
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% This program is free software; you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation; either version 2 of the License, or
% (at your option) any later version.
%
% This program is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You can obtain a copy of the GNU General Public License from
% http://www.gnu.org/copyleft/gpl.html or by writing to
% Free Software Foundation, Inc.,675 Mass Ave, Cambridge, MA 02139, USA.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
persistent w_u m_u v_u w_v m_v v_v dpwtdef
% initialize persistent variables
if ~numel(w_u)
% voiced/unvoiced decision based on 2-element feature vector
% (a) mean power of the frame's log-freq spectrum (normalized so its short-term average is LTASS)
% (b) sum of the power in the first three peaks
%===== VUV
if nargin>3 && any(m=='x')
fxpefac_g; % read in GMM parameters
fxpefac_w; % read in Weights parameters
else
w_u=[0.2123723 0.207788 0.2701817 0.1293616 0.04741722 0.1328791 ]';
m_u=[0.2220388 0.4067706 ;
0.04567656 0.4016914 ;
0.8415278 0.3192158 ;
0.2194808 0.1910079 ;
1.6347 0.5819833 ;
1.181519 0.6996485 ];
v_u=reshape([0.01413822 0.003357913 0.003357913 0.01786169 ;
0.0009377269 0.0006220489 0.0006220489 0.03422057 ;
0.1233703 0.004299293 0.004299293 0.007660504 ;
0.01779449 0.002078821 0.002078821 0.001605052 ;
1.110173 0.00718649 0.00718649 0.005734435 ;
0.5477135 -0.00182316 -0.00182316 0.05659796 ]',[2 2 6]);
w_v=[0.07758689 0.2109879 0.1856225 0.06853158 0.2701563 0.1871148 ]';
m_v=[1.208656 0.3365564 ;
1.216643 0.5971916 ;
4.08585 1.240948 ;
8.322102 1.349939 ;
1.734108 1.168643 ;
0.5107205 0.940308 ];
v_v=reshape([0.06181574 0.002950501 0.002950501 0.004528442 ;
0.2946077 0.01433284 0.01433284 0.02684239 ;
2.508473 -0.03310555 -0.03310555 0.1098579 ;
14.17252 -0.09009174 -0.09009174 0.07989255 ;
0.5834894 -0.07854027 -0.07854027 0.1108958 ;
0.05978017 0.005528601 0.005528601 0.1309329 ]',[2 2 6]);
end
%===== PDP
% dfm = -0.4238; % df mean
% dfv = 3.8968; % df variance (although treated as std dev here)
% delta = 0.15;
% dflpso=[dfm 0.5/(log(10)*dfv^2) -log(2*delta/(dfv*sqrt(2*pi)))/log(10)]; % scale factor & offset for df pdf
% dpwtdef=[1.0000, 0.8250, 1.3064, 1.9863]; % default DP weights
dpwtdef=[1.0000, 0.8250, 0.01868, 0.006773, 98.9, -0.4238]; % default DP weights
%===== END
end
% Algorithm parameter defaults
p.fstep=5; % frequency resolution of initial spectrogram (Hz)
p.fmax=4000; % maximum frequency of initial spectrogram (Hz)
p.fres = 20; % bandwidth of initial spectrogram (Hz)
p.fbanklo = 40; % low frequency limit of log filterbank (Hz)
p.mpsmooth = 201; % width of smoothing filter for mean power
% p.maxtranf = 1000; % maximum value of tranf cost term
p.shortut = 7; % max utterance length to average power of entire utterance
p.pefact = 1.5; % shape factor in PEFAC filter
p.numopt = 3; % number of possible frequencies per frame
p.flim = [60 400]; % range of feasible fundamental frequencies (Hz)
p.w = dpwtdef; % DP weights
% p.rampk = 1.1; % constant for relative-amplitude cost term
% p.rampcz = 100; % relative amplitude cost for missing peak
p.tmf = 2; % median frequency smoothing interval (s)
p.tinc = 0.01; % default frame increment (s)
% update parameters from pp argument
if nargin>=5 && isstruct(pp)
fnq=fieldnames(pp);
for i=1:length(fnq)
if isfield(p,fnq{i})
p.(fnq{i})=pp.(fnq{i});
end
end
end
% Sort out input arguments
if nargin>=3 && numel(tinc)>0
p.tinc = tinc; % 0.01 s between consecutive time frames
end
if nargin<4
m='';
end
% Spectrogram of the mixture
fmin = 0; fstep = p.fstep; fmax = p.fmax;
fres = p.fres; % Frequency resolution (Hz)
[tx,f,MIX]=spgrambw(s,fs,fres,[fmin fstep fmax],[],p.tinc);
nframes=length(tx);
txinc=tx(2)-tx(1); % actual frame increment
% ==== we could combine spgrambw and filtbankm into a single call to spgrambw or use fft directly ====
% Log-frequency scale
[trans,cf]=filtbankm(length(f),2*length(f)-1,2*f(end),p.fbanklo,f(end),'usl');
O = MIX*trans'; % Original spectrum in Log-frequency scale
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Amplitude Compression
% Calculate alpha based on LTASS ratios
ltass = stdspectrum(6,'p',cf);
auxf = [cf(1),(cf(1:end-1)+cf(2:end))./2,cf(end)];
ltass = ltass.*diff(auxf); % weight by bin width
% estimated ltass
O = O.*repmat(diff(auxf),nframes,1); % weight spectrum by bin width
if tx(end)<p.shortut % if it is a short utterance
eltass = mean(O,1); % mean power per each frequency band
eltass = smooth(eltass,p.mpsmooth); % smooth in log frequency
eltass= eltass(:).'; % force a row vector
% Same mean power per frame as ltass
cte = mean(ltass)/mean(eltass);
eltass = eltass.*cte; % normalize to have the same mean as LTASS
O = O.*cte;
% Linear AC
alpha = (ltass)./(eltass);
alpha = alpha(:).';
alpha = repmat(alpha,nframes,1);
O = O.*alpha; % force O to have an average LTASS spectrum
% ==== should perhaps exclude the silent portions ***
else % long utterance
tsmo = 2; % time smoothing over 1 sec
stt = round(tsmo/txinc);
filttime = [ones(stt,1); zeros(stt-1,1)];
filtfreq = ones(1,p.mpsmooth);
eltass = imfilter(O,filttime);
eltass = imfilter(eltass,filtfreq); % filter in time and log frequency
% Same mean power per frame than ltass
cte = repmat(mean(ltass),nframes,1)./mean(eltass,2);
eltass = eltass.*repmat(cte,1,length(cf));
O = O.*repmat(cte,1,length(cf));
% Linear AC
alpha = repmat(ltass,nframes,1)./(eltass);
O = O.*alpha;
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Create the filter to detect the harmonics
ini = find(cf>2*cf(1));
sca = cf/cf(ini(1)); % normalize bin frequencies to start at approximately 0.5
sca = sca(sca<10.5 & sca>0.5); % restrict to 0.5 - 10.5 times fundamental
filh = -log10(p.pefact-cos(2*pi*sca));
filh = filh-mean(filh); % force filter to be zero mean
posit = find(sca>=1); % ==== this should just equal ini(1) ====
if ~mod(length(posit),2)
filh = [filh 0]; % force to be an odd length after central tap
end
negat = find(sca<1);
numz = length(posit)-1-length(negat);
filh = filh./max(filh);
filh = [zeros(1,numz) filh];
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Filter the log-frequency scaled spectrogram
B = imfilter(O,filh); % ==== no good reason to use imfilter here ====
% Feasible frequency range
numopt = p.numopt; % Number of possible fundamental frequencies per frame
flim = p.flim;
pfreq = find(cf>flim(1) & cf<flim(2));
ff = zeros(nframes,numopt);
amp = zeros(nframes,numopt);
for i=1:nframes
[pos,peak]=findpeaks(B(i,pfreq),[],5/(cf(pfreq(2))-cf(pfreq(1)))); % ==== calculate some out of loop ====
if numel(pos)
[peak,ind]=sort(peak,'descend');
pos = pos(ind); % indices of peaks in the B array
posff = cf(pfreq(pos)); % frequencies of peaks
fin = min(numopt,length(posff));
ff(i,1:fin)=posff(1:fin); % save both frequency and amplitudes
amp(i,1:fin)=peak(1:fin);
% else
% ff(i,:)=0; % ==== unnecessary since they start as zeros ====
% amp(i,:)=0;
end
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Probabilitly of the frame of being voiced
% voiced/unvoiced decision based on 2-element feature vector
% (a) mean power of the frame's log-freq spectrum (normalized so its short-term average is LTASS)
% (b) sum of the power in the first three peaks
pow = mean(O,2)*1e-6;
vuvfea = [pow sum(amp,2)./(pow*1e9)];
pru=gaussmixp(vuvfea,m_u,v_u,w_u); % Probability of being unvoiced
prv=gaussmixp(vuvfea,m_v,v_v,w_v); % Probability of being voiced
% pru = exp(pru);
% prv=exp(prv); % Linear probability
% pv = prv./(prv+pru); % ==== better to write pv=(1+exp(pru-prv)).^(-1) ====
pv=(1+exp(pru-prv)).^(-1);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Dynamic programming
% w(1): relative amp, voiced local cost
% w(2): median pitch deviation cost
% w(3): df cost weight
% w(4): max df cost
% w(5): relative amp cost for missing peaks (very high)
% w(6): df mean
w = p.w;
% Relative amplitude
camp = -amp./repmat(max(amp,[],2),1,numopt); % relative amplitude used as cost
camp(amp==0)=w(5); % If no frequency found
% Time interval for the median frequency
tmf = p.tmf; % in sec
inmf = round(tmf/txinc);
%--------------------------------------------------------------------------
% FORWARDS
% Initialize values
cost = zeros(nframes,numopt);
prev = zeros(nframes,numopt);
medfx = zeros(nframes,1);
dffact=2/txinc;
% First time frame
% cost(1,:) = w(1)*ramp(1,:);
cost(1,:) = w(1)*camp(1,:); % only one cost term for first frame
fpos = ff(1:min(inmf,end),1);
mf=median(fpos(pv(1:min(inmf,end))>0.6)); % calculate median frequency of first 2 seconds
if isnan(mf)
mf=median(fpos(pv(1:min(inmf,end))>0.5));
if isnan(mf)
mf=median(fpos(pv(1:min(inmf,end))>0.4));
if isnan(mf)
mf=median(fpos(pv(1:min(inmf,end))>0.3)); % ==== clumsy way of ensuring that we take the best frames ====
if isnan(mf)
mf=0;
end
end
end
end
medfx(1)=mf;
for i=2:nframes % main dynamic programming loop
if i>inmf
fpos = ff(i-inmf:i,1); % fpos is the highest peak in each frame
mf=median(fpos(pv(1:inmf)>0.6)); % find median frequency over past 2 seconds
if isnan(mf)
mf=median(fpos(pv(1:inmf)>0.5));
if isnan(mf)
mf=median(fpos(pv(1:inmf)>0.4));
if isnan(mf)
mf=median(fpos(pv(1:inmf)>0.3));% ==== clumsy way of ensuring that we take the best frames ====
if isnan(mf)
mf=0;
end
end
end
end
end
medfx(i)=mf;
% Frequency difference between candidates and cost
df = dffact*(repmat(ff(i,:).',1,numopt) - repmat(ff(i-1,:),numopt,1))./(repmat(ff(i,:).',1,numopt) + repmat(ff(i-1,:),numopt,1));
costdf=w(3)*min((df-w(6)).^2,w(4));
% Cost related to the median pitch
if mf==0 % this test was inverted in the original version
costf = zeros(1,numopt);
else
costf = abs(ff(i,:) - mf)./mf;
end
[cost(i,:),prev(i,:)]=min(costdf + repmat(cost(i-1,:),numopt,1),[],2); % ==== should we allow the possibility of skipping frames ? ====
cost(i,:)=cost(i,:)+w(2)*costf + w(1)*camp(i,:); % add on costs that are independent of previous path
end
% Traceback
fx=zeros(nframes,1);
best = zeros(nframes,1);
nose=find(cost(end,:)==min(cost(end,:))); % ==== bad method (dangerous) ===
best(end)=nose(1);
% ff = [ff zeros(nframes,1)]; % not clear why this was here
fx(end)=ff(end,best(end));
for i=nframes:-1:2
best(i-1)=prev(i,best(i));
fx(i-1)=ff(i-1,best(i-1));
end
if nargout>=4
fv.vuvfea=vuvfea; % voiced-unvoiced features
fv.best=best; % selected path
fv.ff=ff; % pitch candidates
fv.amp=amp; % pitch candidate amplitudes
fv.medfx=medfx; % median pitch
fv.w=w; % DP weights
fv.dffact=dffact; % df scale factor
end
if ~nargout || any(m=='g') || any(m=='G')
nax=0; % number of axes sets to link
msk=pv>0.5; % find voiced frames as a mask
fxg=fx;
fxg(~msk)=NaN; % allow only good frames
fxb=fx;
fxb(msk)=NaN; % allow only bad frames
if any(m=='G') || ~nargout && ~any(m=='g')
clf;
spgrambw(s,fs,'ilcwpf'); % draw spectrogram with log axes
hold on
plot(tx,log10(fxg),'-b',tx,log10(fxb),'-r'); % fx track
yy=get(gca,'ylim');
plot(tx,yy(1)+yy*[-1;1]*(0.02+0.05*pv),'-k'); % P(V) track
hold off
nax=nax+1;
axh(nax)=gca;
if any(m=='g')
figure; % need a new figure if plotting two graphs
end
end
if any(m=='g')
ns=length(s);
[tsr,ix]=sort([(1:ns)/fs 0.5*(tx(1:end-1)+tx(2:end))']); % intermingle speech and frame boundaries
jx(ix)=1:length(ix); % create inverse index
sp2fr=jx(1:ns)-(0:ns-1); % speech sample to frame number
spmsk=msk(sp2fr); % speech sample voiced mask
sg=s;
sg(~spmsk)=NaN; % good speech samples only
sb=s;
sb(spmsk)=NaN; % bad speech samples only
clf;
subplot(5,1,1);
plot(tx,pv,'-b',(1:ns)/fs,0.5*mod(cumsum(fx(sp2fr)/fs),1)-0.6,'-b');
nax=nax+1;
axh(nax)=gca;
ylabel('\phi(t), P(V)');
set(gca,'ylim',[-0.65 1.05]);
subplot(5,1,2:3);
plot((1:ns)/fs,sg,'-b',(1:ns)/fs,sb,'-r');
nax=nax+1;
axh(nax)=gca;
subplot(5,1,4:5);
plot(tx,fxg,'-b',tx,fxb,'-r');
ylabel('Pitch (Hz)');
% semilogy(tx,fxg,'-b',tx,fxb,'-r');
% ylabel(['Pitch (' yticksi 'Hz)']);
set(gca,'ylim',[min(fxg)-30 max(fxg)+30]);
nax=nax+1;
axh(nax)=gca;
end
if nax>1
linkaxes(axh,'x');
end
end
function y=smooth(x,n)
nx=length(x);
c=cumsum(x);
y=[c(1:2:n)./(1:2:n) (c(n+1:end)-c(1:end-n))/n (c(end)-c(end-n+2:2:end-1))./(n-2:-2:1)];