% Compute the utterance-based recognition rates for speaker identification.
% To run this program, you need to run goFrameRrWrtGaussianNum.m first to get gmmParamWrtGaussianNum.mat

% Roger Jang, 20070531, 20090131

tInit=clock;
addMyPath;
sidOpt=sidOptSet;
eval(sprintf('load %s/speakerSet.mat', sidOpt.outputDir));
if sidOpt.maxSpeakerNum<length(speakerSet1)
	speakerSet1(sidOpt.maxSpeakerNum+1:end)=[];
	speakerSet2(sidOpt.maxSpeakerNum+1:end)=[];
end
eval(sprintf('load %s/gmmParamWrtGaussianNum.mat', sidOpt.outputDir));
caseNum=length(gmmData);

if ~exist([sidOpt.outputDir, '/utterRrVec.mat'])
	% ====== Feature transformation if necessary
	if sidOpt.feaTransformMode
		speakerSet1=speakerSetFeaTransform(speakerSet1, transformVec);	% transformVec is obtained in goFrameRrWrtGaussianNum.m
		speakerSet2=speakerSetFeaTransform(speakerSet2, transformVec);	% transformVec is obtained in goFrameRrWrtGaussianNum.m
	end
	% ====== Inside/outputside test	
	for i=1:caseNum
		fprintf('%d/%d: No. of Gaussians = %d ===> ', i, caseNum, length(gmmData(i).gmm));
		utterRrVec1(i)=speakerIdentify(speakerSet1, gmmData(i).gmm, sidOpt.useIntGmm);	% Inside test
		utterRrVec2(i)=speakerIdentify(speakerSet2, gmmData(i).gmm, sidOpt.useIntGmm);	% Outside test
		fprintf('inside RR = %.2f%%, outside RR = %.2f%%\n', utterRrVec1(i)*100, utterRrVec2(i)*100);
	end
	eval(sprintf('save %s/utterRrVec utterRrVec1 utterRrVec2', sidOpt.outputDir));
else
	fprintf('Loading %s/utterRrVec.mat...\n', sidOpt.outputDir);
	eval(sprintf('load %s/utterRrVec.mat', sidOpt.outputDir));
end

for i=1:caseNum
	gaussianNum(i)=length(gmmData(i).gmm(1).gmmParam);
end

[maxRr, gaussianNumBestIndex]=max(utterRrVec2);
gaussianNumBest=gaussianNum(gaussianNumBestIndex);			% This is to be used in the following script
subplot(2,1,2);
plot(1:caseNum, utterRrVec1*100, 'square-', 1:caseNum, utterRrVec2*100, 'o-');
axis([-inf inf 0 100]);
line(gaussianNumBestIndex, 100*utterRrVec2(gaussianNumBestIndex), 'marker', '*', 'color', 'r', 'linestyle', 'none');
title(sprintf('Utterance-based recognition rates (Best=%g%% when gaussianNum=%d)', 100*utterRrVec2(gaussianNumBestIndex), gaussianNumBest));
set(gca, 'xtick', 1:length(gaussianNum));	% Avoid xtick=[1, 1.5, 2, 2.5, ...]
set(gca, 'xticklabel', gaussianNum);
xlabel('No. of Gaussians components');
ylabel('Recognition rates (%)');
legend('Inside-test recog. rate', 'Outside-test recog. rate', 'Best setting', 4);
grid on;
eval(sprintf('print -dpng %s/%s', sidOpt.outputDir, mfilename));
fprintf('%s ===> %g seconds\n', mfilename, etime(clock, tInit));