Benedikt Loesch
Chair of System Theory and Signal Processing, University of Stutgart
benedikt.loesch@lss.uni-stuttgart.de

My algorithm first performs a segmentation of the recording into segments 
where 2 sources are active and then performs frequency domain ICA on each 
segment. Permutation is solved using the State Coherence Transform.
The approach is described in detail in a paper submitted to LVA/ICA 2010:
[1] B. Loesch, and B. Yang: "Adaptive Segmentation and Separation of 
Determined Convolutive Mixtures under Dynamic Conditions", submitted to 
LVA/ICA 2010

The running time of the MATLAB implementation for each recording of length 173 
seconds is about 14 minutes on a single core of Intel(R) Xeon(R) CPU E5440 @ 
2.83GHz.
This can be improved to about 4 minutes by using a C implementation.
The cited running times are for 25 iterations of the ICA algorithm.
The running time can be decreased further by decreasing the number of 
iterations. The performance for the example dataset as a function of the 
number of iterations is given below:
iterations	   SIR without A weighting            SIR with A-weighting
10			    9.2204					11.5490
15			    9.2357 					11.9432
20			    9.3596					12.2526
25			    9.2522				        12.3271

Reducing the number of iterations to 10 or 15 leads to a slight decrease in 
SIR performance, but the running time can be decreased by a factor of approx 
2.