% EXAMPLE_CONV Example of use of software routines for the separation of a
% convolutive mixture with 5cm microphone spacing
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Copyright 2008 Emmanuel Vincent
% This software is distributed under the terms of the GNU Public License
% version 3 (http://www.gnu.org/licenses/gpl.txt)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Input time-frequency representation
x=wavread('dev1/dev1_nodrums_synthconv_250ms_5cm_mix.wav').';
X=stft_multi(x);

% Manual setting of the number of sources
nsrc=3;

% Estimation of the frequency-dependent mixing matrix
wlen=2*(size(X,1)-1);
Ae=estmix_conv_5cm([X;conj(X(wlen/2:-1:2,:,:))],nsrc);
Ae=Ae(:,:,1:wlen/2+1);

% Evaluation of the estimated frequency-dependent mixing matrix
load('dev1/dev1_nodrums_synthconv_250ms_5cm_filt.mat');
A=sum(reshape(cat(3,A,zeros(2,nsrc,wlen*ceil(size(A,3)/wlen)-size(A,3))),2,nsrc,wlen,ceil(size(A,3)/wlen)),4);
A=fft(A,[],3);
A=A(:,:,1:wlen/2+1);
[MER,perm]=bss_eval_mix(Ae,A);
mkdir('results');
save('results/dev1_nodrums_synthconv_250ms_5cm_filt.mat','Ae');

% Source separation via binary masking
S=sep_binmask(X,Ae);
se=istft_multi(S,160000);
for j=1:nsrc,
    wavwrite(se(j,:).',16000,['results/dev1_nodrums_synthconv_250ms_5cm_src_' int2str(j) '.wav']);
end

% Evaluation of the estimated sources
s=zeros(3,160000);
for j=1:3,
    s(j,:)=wavread(['dev1/dev1_nodrums_src_' int2str(j) '.wav'],[16001 176000]).';
end
[SDRs,SIRs,SARs,perms]=bss_eval_sources(se,s);

% Computation of the spatial source images
I=src_image(S,Ae);
ie=istft_multi(I,160000);
for j=1:nsrc,
    wavwrite(reshape(ie(j,:,:),160000,2),16000,['results/dev1_nodrums_synthconv_250ms_5cm_sim_' int2str(j) '.wav']);
end

% Evaluation of the estimated source images
i=zeros(3,160000,2);
for j=1:3,
    i(j,:,:)=reshape(wavread(['dev1/dev1_nodrums_synthconv_250ms_5cm_sim_' int2str(j) '.wav']),1,160000,2);
end
[SDRi,ISRi,SIRi,SARi,permi]=bss_eval_images(ie,i);

% Ideal binary masking benchmark (STFT)
b1=sep_ibm(x,i);
[SDRb1,ISRb1,SIRb1,SARb1]=bss_eval_images_nosort(b1,i);

% Ideal binary masking benchmark (Cochleagram)
b2=zeros(4,160000,2);
for j=1:4,
    ibm=generateIBM(reshape(i(j,:,:),160000,2),reshape(sum(i([1:j-1 j+1:4],:,:),1),160000,2));
    b2(j,:,:)=generateEstimate(x.',ibm);
end
[SDRb2,ISRb2,SIRb2,SARb2]=bss_eval_images_nosort(b2,i);