Mauch MIREX 2010

function df = get_dfs(x,p)

% function to calculate the following onset detection functions

% df{1} = complex spectral difference

% df{2} = spectral difference

% df{3} = phase deviation

x = mean(x,2);

% onset analysis step increment

o_step = p.winlen*2; % should be 1024

% onset analysis winlen

o_winlen = o_step*2; % should be 2048

hlfwin = o_winlen/2; % will be half fft size

% formulate hanningz window function

win = hanningz(o_winlen);

%  win = tukeywin(o_winlen,0.5);

% loop parameters

N = length(x);

pin = 0;

pend = N - o_winlen;

% vectors to store phase and magnitude calculations

theta1 = zeros(hlfwin,1);

theta2 = zeros(hlfwin,1);

oldmag = zeros(hlfwin,1);

% output onset detection functions

df = {};

% df sample number

k = 0;

while pin<pend

    k=k+1;

    % calculate windowed fft frame

    segment = x(pin+1:pin+o_winlen);

    X = fft(fftshift(win.*segment));

    %X = X.*fftshift(hanningz(length(X)));

    X = X(floor(length(X)/2)+1:length(X),:);

  %  X = X.*[length(X):-1:1]';

    % separate into magnitude

    mag = (abs(X));

%  	mag = [0 abs(diff(log(eps+mag)))']';

    % and phase

    theta = angle(X);

    % complexsd part

        dev=princarg(theta-2*theta1+theta2);

%  		keyboard

        meas=oldmag - (mag.*exp(i.*dev));

        df{1}(k) = sum(sqrt((real(meas)).^2+(imag(meas)).^2));

%      % spectral difference part    

%          dev2=sqrt(mag.^2-oldmag.^2);

%          df{2}(k) = sum(abs(dev2(mag>0.1)));

%      

%      

%      % phase deviation part

%          df{3}(k) = sum(abs(dev((mag>0.1))));

%  %        df{3}(k) = sum(abs(dev(end-5:end)));

%          

%      % energy part

%          df{4}(k) = sum(abs(segment));

%          

%      % hfc part    

%          df{5}(k) = sum(mag.*(length(mag):-1:1)');

%      %    df{5}(k) = sum(mag.*rand(hlfwin,1));

%          

    % update vectors

    theta2 = theta1;

    theta1 = theta;

    oldmag = mag;

    pin = pin+o_step;

end

% now interpolate each detection function by a factor of 2,

% to get resolution of 11.6ms

for j=1:1,%5

    df{j} = interp(df{j},2);

    % also check there are no negative elements

    df{j}(df{j}<0) = 0;

    % and scale to sum to unity

%    df{j} = (df{j}+eps)/sum(df{j}+eps);

end