view baseline.m @ 0:acfea2266c6d tip

Baseline classification system. Note that needs ma toolbox (for comoputation of mfccs) and pmtk3 toolbox

function tesCat = baseline(train_files,test_files,out_name)
% Baseline ASC function using mfccs and gmms.
% Input: 
%   - train_files: text file containing the path to the training files
%   - test_files: text file containing the path to the test files
%   - out_name: text file where to write the results
% Output:
%   - est_cat: cell of strings containing the estimated classes of the test
%   files
% 
%% Input parsing and parameters
% Open files, retrieve paths and training categories.
trainData = textscan(fopen(train_files,'r'),'%s');
fclose('all');
nTraFil   = length(trainData{1})/2;
traPath    = cell(nTraFil,1);
traCat    = cell(nTraFil,1);
for i=1:nTraFil
    traPath{i} = trainData{1}{2*i-1};
    traCat{i} = trainData{1}{2*i};
    
    traPath{i} = regexprep(traPath{i},'import','Volumes'); 
end

testData = textscan(fopen(test_files,'r'),'%s');
fclose('all');
nTesFil   = length(testData{1});
tesPath    = cell(nTesFil,1);
tesCat    = cell(nTesFil,1);
for i=1:nTesFil
    tesPath{i} = testData{1}{i};
    tesPath{i} = regexprep(tesPath{i},'import','Volumes'); 
end

% Set parameters
feaNam = {'mfcc'};                                  %compute mfccs
par = struct('fs', 22050,...                        %sampling frequency
             'fft_size', 1024,...                   %size of Fourier transform
             'hopsize', 512,...                     %overlap of Fourier transfrom windows
             'usePreEmphasis', false,...            %do not filter sound prior to feature extraction
             'feaNam',feaNam,...                    %extract mfcc coefficients
             'nGaussCom',5,...                      %use 5 Gaussian mixture components
             'empar', struct('nrandomRestarts',10,...%restart algoritms 5 times and pick the best model
                             'maxIter',50,...       %maximum number of iterations
                             'verbose',false));

%% Training phase
% Extract features from each training signal
traFea = cell(nTraFil,1);
for iFil=1:length(traPath)
    fprintf('\n Extracting %s features from file %s - (%s)...',feaNam{:},traPath{iFil},traCat{iFil});
    traFea{iFil} = ComputeFeatures(traPath{iFil},par);
    fprintf(' done!');
end

% Pre-process features by subtracting mean and dividing by standard
% deviation
muFea = mean([traFea{:}],2);
sdFea    = std([traFea{:}],[],2);
for iFil=1:nTraFil
    traFea{iFil} = (traFea{iFil}-repmat(muFea,1,size(traFea{iFil},2)))./repmat(sdFea,1,size(traFea{iFil},2));
end

% Compute statistical distribution of the featuers belonging to a given
% category using GMMs
staMod = 'GMMs';                %statistical model
uniCat = unique(traCat);        %unique categories
nUniCat = length(uniCat);
models = cell(nUniCat,1);
for iUniCat=1:nUniCat
    fprintf('\n Computing statistical distribution of features from class %s using %s ...',uniCat{iUniCat},staMod);
    idx = strcmpi(traCat,uniCat(iUniCat));    %indexes of trainig files belonging to iUniCat category
    traFeaMat = [traFea{idx}]';
    models{iUniCat} = mixGaussFit(traFeaMat,par.nGaussCom,par.empar);
    fprintf(' done!');
end


%% Test phase
% Repeat processing chain and compute the features extracted from test files
for iFil=1:nTesFil
    fprintf('\n Extracting %s features from file %s...',feaNam{:},tesPath{iFil});
    tesFea = ComputeFeatures(tesPath{iFil},par);
    % pre-process features using the mean and standard deviation values
    % computed on the training set
    tesFea = (tesFea-repmat(muFea,1,size(tesFea,2)))./repmat(sdFea,1,size(tesFea,2));
    fprintf(' done!');
    % Compute likelihood scores 
    scores = nan(length(models),1);
    for iModel=1:length(models)
        scores(iModel) = sum(mixGaussLogprob(models{iModel},tesFea'));   %returns the negative loglikelihood of the data given the model iModel
    end
    [~, ind] = max(scores);
    tesCat{iFil} = uniCat{ind(1)};
end
filNam = [mfilename,'_',datestr(now,'ddmmyyyy-HH:MM'),'.txt'];
fId = fopen(['results', filesep, out_name],'w');
for iFil=1:nTesFil
    fprintf(fId,[tesPath{iFil} '\t']);
    fprintf(fId,[tesCat{iFil} '\n']);
end
fclose(fId);