Mauch MIREX 2010

function [beats,localscore,timesig,ppath] = newtt_hmm_dp(df,tempo)

if nargin<2

	tempo = 0;

end

beats = [];

p = bt_parms;

df = df(:)';

%  

% df = df/alpha_norm(9,df);

df = qnormalise(df,9,2);

% work out period track with viterbi decoding

[ppath,obs,prob,wv]= tempo_hmm(df,p.step,tempo);

%  ppp=textread('home/studio/Desktop/path.txt');

%  ppp = [ppp(:)' ppp(end)*ones(1,2)];

%  ppath = ppp;

ppath(ppath<20) = 20;

dp = abs(diff(ppath));

dp = [0 dp(:)'];

dp(dp>16) = 16; 

%  plot(dp),pause(0.01);

% now incorporate ellis' dynamic programming,

% but for time-varying period....! 

tempmat = (ppath'*ones(1,p.step))';

pd = round(tempmat(:));

%  length(pd)

%  df = abs(diff(df));

% prob = 0.5+0.5*prob/max(prob);

% prob = 4*prob;

%  prob = 4*prob/max(prob);

%  a=ftacf(adapt_thresh(df));

%  a=a(1:512);

%  [tmp,r]=getperiod(a,wv,0,128,1,0);

%  r=sunity(r);

%  val=max(r);

%  tightness = 100*val;

tightness= [4]; % want to try and get a tightness out of the signal automatically...

alpha=0.8;

mpd = round(median(ppath));

templt = exp(-0.5*(([-mpd:mpd]/(mpd/32)).^2));

localscore = conv(templt,df);

% this removes the offset imposed by the filtering...

localscore = localscore(round(length(templt)/2)+[1:length(df)]);

localscore = localscore/sum(localscore);

aacf = ftacf(localscore);

[timesig] = findmeter_mm(aacf(1:512),mpd);

%  %  subplot(311); plot(localscore);

%  %  al = localscore;

%  %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%  %%%% little bootstrap on phase

%  %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%  apin = 0;

%  awinlen=512;

%  astep = 128;

%  apend = length(localscore)-2*awinlen;

%  ak = 0;

%  while (apin<apend)

%  	ak = ak+1;

%  	align = getalignment2(localscore(apin+1:apin+awinlen),ones(1,pd(ak)),pd(ak),0);

%  	try

%  		localscore(apin+align-4:apin+align+4) = (1+gausswin(9)').*localscore(apin+align-4:apin+align+4);

%  	catch

%  		temp_a = 1;

%  	end	

%  	apin=apin+astep;

%  end

%  

%  %  subplot(312); plot(localscore);

%  %  subplot(313); plot(al-localscore);

backlink = zeros(1,length(localscore));

cumscore = zeros(1,length(localscore));

starting = 1;

for i = 1:length(localscore)

%  	pause;

%  	i

% search range for previous beat

	prange = round(-2*pd(i)):-round(pd(i))/2;

%   want to update tightness based on periodicity change..

	% or as andrew does.. whether it's an important beat.

	% Skewed window

	mu = -pd(i);

%  	tightness = prob(ceil(i/128));

%  	sigma = pd(i)/4;

%  	txwt = exp(-0.5 * ((prange - mu)./sigma).^2) ./ (sqrt(2*pi) .* sigma);

	txwt = exp( -0.5*  (  (  ( tightness    - 1* log(1+dp(ceil(i/128))) )   *   log(prange/-pd(i))).^2)  );

%  	txwt = -(prange ./ mu.^ 2) .* exp(-prange .^ 2 ./ (2*mu .^ 2));

%  	txwt = txwt.^2;

%  	keyboard

	timerange = i + prange;

% Are we reaching back before time zero?

	zpad = max(0, min(1-timerange(1),length(prange)));

% Search over all possible predecessors and apply transition 

% weighting

	scorecands = txwt .* [zeros(1,zpad),cumscore(timerange(zpad+1:end))];

% Find best predecessor beat

	[vv,xx] = max(scorecands);

% Add on local score

	cumscore(i) = alpha*vv + (1-alpha)*localscore(i);

% special case to catch first onset

%  if starting == 1 & localscore(i) > 100*abs(vv)

	if starting == 1 & localscore(i) < 0.01*max(localscore);

		backlink(i) = -1;

	else

		backlink(i) = timerange(xx);

		% prevent it from resetting, even through a stretch of silence

		starting = 0;

	end

end

%try

%	align = getalignment2(localscore(end-1024:end),ones(1,2*pd(end)),2*pd(end),0);

%catch

	align = getalignment2(localscore(end-512:end),ones(1,1*pd(end)),1*pd(end),0);

%end

b = [];

b = length(df) - align;

%  b = bestendx;

while backlink(b(end)) > 0

b = [b,backlink(b(end))];

end

b = fliplr(b);

% return beat times in secs

beats = (b*512); % seems to be an offset.

beats(beats<1) = [];

%  

%  otherbeats{kk} = b;

%  score(kk) = dfscore(b,df);		

%  [val,ind] = max(score);

%  beats = otherbeats{ind};

%  beats = 512*(length(df)-beats/512); 

%  beats = sort(beats);