Chris@2: function [ph pz sumY] = transcriptionMultipleTemplates(filename,iter,sz,su)
Chris@2: 
Chris@2: 
Chris@2: % Load note templates
Chris@2: load('noteTemplatesBassoon');      W(:,:,1) = noteTemplatesBassoon;
Chris@2: load('noteTemplatesCello');        W(:,:,2) = noteTemplatesCello;
Chris@2: load('noteTemplatesClarinet');     W(:,:,3) = noteTemplatesClarinet;
Chris@2: load('noteTemplatesFlute');        W(:,:,4) = noteTemplatesFlute;
Chris@2: load('noteTemplatesGuitar');       W(:,:,5) = noteTemplatesGuitar;
Chris@2: load('noteTemplatesHorn');         W(:,:,6) = noteTemplatesHorn;
Chris@2: load('noteTemplatesOboe');         W(:,:,7) = noteTemplatesOboe;
Chris@2: load('noteTemplatesTenorSax');     W(:,:,8) = noteTemplatesTenorSax;
Chris@2: load('noteTemplatesViolin');       W(:,:,9) = noteTemplatesViolin;
Chris@9: 
Chris@9: %% SptkBGCl -> piano (it stands for Sampletek Steinway "Black Grand").
Chris@9: %% It took me a while to figure this out! (It is documented in the
Chris@9: %% MAPS database)
Chris@2: load('noteTemplatesSptkBGCl');     W(:,:,10) = noteTemplatesSptkBGCl;
Chris@2: 
Chris@2: %pitchActivity = [14 16 30 40 20 21 38 24 35 1; 52 61 69 76 56 57 71 55 80 88]';
Chris@2: pitchActivity = [16 16 30 40 20 21 38 24 35 16; 52 61 69 73 56 57 71 55 73 73]';
Chris@2: 
Chris@2: 
Chris@5: %% this turns W0 into a 10x88 cell array in which W0{instrument}{note}
Chris@5: %% is the 545x1 template for the given instrument and note number.
Chris@2: W = permute(W,[2 1 3]);
Chris@2: W0 = squeeze(num2cell(W,1))';
Chris@5: 
Chris@2: clear('noteTemplatesBassoon','noteTemplatesCello','noteTemplatesClarinet','noteTemplatesFlute','noteTemplatesGuitar',...
Chris@2:     'noteTemplatesHorn','noteTemplatesOboe','noteTemplatesTenorSax','noteTemplatesViolin','noteTemplatesSptkBGCl','W');
Chris@2: 
Chris@2: 
Chris@2: % Compute CQT
Chris@5: 
Chris@5: %% The CQT parameters are hardcoded in computeCQT. It has frequency
Chris@5: %% range 27.5 -> samplerate/3, 60 bins per octave, a 'q' of 0.8 (lower
Chris@5: %% than the maximum, and default, value of 1), 'atomHopFactor' 0.3
Chris@5: %% rather than the default 0.25 (why?), Hann window, default sparsity
Chris@10: %% threshold. The CQT obtained is the interpolated real-valued
Chris@10: %% magnitude spectrogram rather than the complex output.
Chris@10: 
Chris@10: %% The audio is always resampled to 44100Hz (if it isn't at that rate
Chris@10: %% already) and mixed down to mono.
Chris@10: 
Chris@10: %% The computed CQT parameters actually obtained are:
Chris@10: %%   10 octaves
Chris@10: %%   highest frequency 14700Hz
Chris@10: %%   lowest frequency 14.5223Hz
Chris@10: %%   column height 600 (60 bpo * 10 oct)
Chris@10: %% But only bins 56:600 are used, the first 55 are dropped, leaving
Chris@10: %% 545 bins per column. I *think* the spectrogram is the "right" way
Chris@10: %% up at this point so those first 55 bins are the lowest-frequency
Chris@10: %% ones, meaning the frequency range actually returned is 27.4144Hz
Chris@10: %% to 14700Hz.
Chris@5: 
Chris@5: %% for a 43.5 second 44.1 KHz audio file, intCQT will be a 545x30941
Chris@5: %% array, one column every 0.0014 seconds.
Chris@2: [intCQT] = computeCQT(filename);
Chris@5: 
Chris@5: %% X is sampled from intCQT at 7.1128-column intervals, giving
Chris@5: %% 4350x545 in this case, so clearly 100 columns per second; then
Chris@5: %% transposed
Chris@2: X = intCQT(:,round(1:7.1128:size(intCQT,2)))';
Chris@5: 
Chris@48: %% median filter to remove broadband noise (i.e we filter across
Chris@48: %% frequency rather than time)
Chris@2: noiseLevel1 = medfilt1(X',40);
Chris@2: noiseLevel2 = medfilt1(min(X',noiseLevel1),40);
Chris@2: X = max(X-noiseLevel2',0);
Chris@5: 
Chris@5: %% take every 4th row. We had 100 per second (10ms) so this is 40ms as
Chris@48: %% the comment says. It's not clear to me why we denoise before doing
Chris@48: %% this rather than after? Y is now 1088x545 in our example and looks
Chris@48: %% pretty clean as a contour plot.
Chris@2: Y = X(1:4:size(X,1),:);  % 40ms step
Chris@5: 
Chris@5: %% a 1x1088 array containing the sum of each column. Doesn't appear to
Chris@8: %% be used in here, but it is returned to the caller.
Chris@2: sumY = sum(Y');
Chris@5: 
Chris@2: clear('intCQT','X','noiseLevel1','noiseLevel2');
Chris@2: 
Chris@2: fprintf('%s','done');
Chris@2: fprintf('\n');
Chris@2: fprintf('%s',['Estimating F0s...........']);
Chris@2: 
Chris@2: % For each 2sec segment, perform SIPLCA with fixed W0
Chris@2: ph = zeros(440,size(Y,1));
Chris@2: pz = zeros(88,size(Y,1));
Chris@2: 
Chris@2: for j=1:floor(size(Y,1)/100)
Chris@2: 
Chris@2:     x=[zeros(2,100); Y(1+(j-1)*100:j*100,:)'; zeros(2,100)];
Chris@2:     [w,h,z,u,xa] = cplcaMT( x, 88, [545 1], 10, W0, [], [], [], iter, 1, 1, sz, su, 0, 1, 1, 1, pitchActivity);
Chris@2:    
Chris@2:     H=[]; for i=1:88 H=[H; h{i}]; end;
Chris@2:     ph(:,1+(j-1)*100:j*100) = H;
Chris@2:     Z=[]; for i=1:88 Z=[Z z{i}]; end;
Chris@2:     pz(:,1+(j-1)*100:j*100) = Z';   
Chris@2:     perc = 100*(j/(floor(size(Y,1)/100)+1));
Chris@2:     fprintf('\n');
Chris@2:     fprintf('%.2f%% complete',perc);
Chris@2: end;
Chris@2: 
Chris@2: len=size(Y,1)-j*100; % Final segment
Chris@2: 
Chris@2: if (len >0)
Chris@2: x=[zeros(2,len); Y(1+j*100:end,:)'; zeros(2,len)];
Chris@2: [w,h,z,u,xa] = cplcaMT( x, 88, [545 1], 10, W0, [], [], [], iter, 1, 1, sz, su, 0, 1, 1, 1, pitchActivity);
Chris@2: fprintf('\n');
Chris@2: fprintf('100%% complete');
Chris@2: 
Chris@2: H=[]; for i=1:88 H=[H; h{i}]; end;
Chris@2: ph(:,1+j*100:end) = H;
Chris@2: Z=[]; for i=1:88 Z=[Z z{i}]; end;
Chris@2: pz(:,1+j*100:end) = Z'; 
Chris@5: end;