29 % (We used stochastic gradient descent, where each value in the 30 % array was used as the learning rate for one pass through the data.) 31 % Note: If this doesn't work for you, feel free to fiddle with learning 32 % rates, etc.

( ones ( size ( mix, 1 ), 1 ) * max ( abs ( mix ) ) ) 9 10 % handle writing in both matlab and octave 11 v = version 12 if ( v ( 1 ) < = ' 3' ) % assume this is octave 13 wavwrite ( ' mix1.wav', normalizedMix ( :, 1 ), Fs, 16 ) 14 wavwrite ( ' mix2.wav', normalizedMix ( :, 2 ), Fs, 16 ) 15 wavwrite ( ' mix3.wav', normalizedMix ( :, 3 ), Fs, 16 ) 16 wavwrite ( ' mix4.wav', normalizedMix ( :, 4 ), Fs, 16 ) 17 wavwrite ( ' mix5.wav', normalizedMix ( :, 5 ), Fs, 16 ) 18 else 19 wavwrite ( normalizedMix ( :, 1 ), Fs, 16, ' mix1.wav' ) 20 wavwrite ( normalizedMix ( :, 2 ), Fs, 16, ' mix2.wav' ) 21 wavwrite ( normalizedMix ( :, 3 ), Fs, 16, ' mix3.wav' ) 22 wavwrite ( normalizedMix ( :, 4 ), Fs, 16, ' mix4.wav' ) 23 wavwrite ( normalizedMix ( :, 5 ), Fs, 16, ' mix5.wav' ) 24 end 25 26 W = eye ( 5 ) % initialize unmixing matrix 27 28 % this is the annealing schedule I used for the learning rate. dat % load mixed sources 5 Fs = 11025 %sampling frequency being used 6 7 % listen to the mixed sources 8 normalizedMix = 0.99 * mix.

Fs 16 ps4