----------------------------------------------------------------
Contact information:
Dang Hai Tran Vu,
Department of Communications Engineering,
University of Paderborn, Germany

Algorithm overview:
Source separation is perfomed by EM algorithm using complex Watson
distribution in frequency domain (see [1]). Compared to [1] an
additional component pdf was introduced in the mixture pdf to model
noise-only time-frequency slots. Permutation alignment is performed by
finding mappings which minimize the inter frequency correlation among
the posteriors of different outputs. Linear MVDR beamforming with TDOA
based gain normalization was employed for spatial separation and noise
suppression. Finally, a single channel non-linear postfilter based on
spectral subtraction is applied where the noise and interferers spectrum
estimation is controlled by the posteriors. 
The music source was considered as a third source in the algorithm. Hence, there no special treatment regarding the music source.

Remarks:
Although the music extraction is not included in this task i have included the output of the BSS algorithm for comparison.
File naming rule is:
Source 1 -> Male Speaker
Source 2 -> Female Speaker
Source 3 -> Music Source

Average running time per file of MATLAB implementation on a Intel Core
i7 920@2.67GHz: 88s


Bibliographical reference:
[1] Dang Hai Tran Vu and Reinhold Haeb-Umbach, "BLIND SPEECH SEPARATION
EMPLOYING DIRECTIONAL STATISTICS IN AN EXPECTATION MAXIMIZATION
FRAMEWORK", in IEEE Proc. ICASSP2010, 2010.

-----------------------------------------------------------------