function mixer = gfb_mixer_new(analyzer, delay, iterations)
%GFB_MIXER_NEW Create new mixer
% Usage: mixer = gfb_mixer_new(analyzer, delay, iterations)
%
% Input parameters:
% analyzer : A gfb_analyzer structure as created by gfb_analyzer_new. The
% mixer created by this function can act as part of a synthesizer
% that resynthesizes the output of this analyzer
% delay : A gfb_Delay structure as created by gfb_delay_new, Together with
% the mixer created by this function, this delay can form a
% synthesizer that resynthesizes the output of the analyzer
% iterations : The gain factors are approximated numerically in iterations.
% If this parameter is omitted, then the number of iterations will
% be GFB_GAINCALC_ITERATIONS (see gfb_set_constants.m, usually
% =100)
%
% `gfb_mixer_new` creates a `gfb_mixer` object with gain factors suitable
% to calculate a weighted sum of the bands present in the output of the
% given delay. The gain factors are computed using a numerical optimization
% method described in [Herzke & Hohmann 2007].
% author : tp
% date : Jan 2002, Nov 2003, Mar & Nov 2006, Jan Feb 2007
global GFB_GAINCALC_ITERATIONS;
gfb_set_constants;
mixer.type = 'gfb_mixer';
center_frequencies = analyzer.center_frequencies_hz;
number_of_bands = length(center_frequencies);
sampling_frequency = analyzer.fs;
% The center frequencies in the z plain
z_c = exp(2i * pi * center_frequencies(:) / sampling_frequency);
mixer.gains = ones(number_of_bands, 1);
% compute the frequency response of each filter (col) at the center
% frequencies of all filters (row)
pos_f_response = ...
gfb_analyzer_zresponse(analyzer, z_c);
neg_f_response = ...
gfb_analyzer_zresponse(analyzer, conj(z_c));
% apply delay and phase correction
for band = [1:number_of_bands]
pos_f_response(:,band) = pos_f_response(:,band) * ...
delay.phase_factors(band) .* ...
z_c .^ -delay.delays_samples(band);
neg_f_response(:,band) = neg_f_response(:,band) * ...
delay.phase_factors(band) .* ...
conj(z_c) .^ -delay.delays_samples(band);
end
% combine responses at positive and negative responses to yield
% responses for real part.
f_response = (pos_f_response + conj(neg_f_response)) / 2;
if (nargin < 4)
iterations = GFB_GAINCALC_ITERATIONS;
end
for i = [1:iterations]
% add selected spectrum of all bands with gain factors
selected_spectrum = f_response * mixer.gains;
% calculate better gain factors from result
mixer.gains = mixer.gains ./ abs(selected_spectrum);
end
mixer.gains = mixer.gains.';