Both datasets were processed with the on-line Semi-blind source Extraction (SBSE) algorithm described in [1].
The algorithm is coded in C and works in real-time on an Intel CPU Xeon E5520  @ 2.27GHz

The initial prior W_{prior} is estimated with a batch ICA based on the weighted scaled Natural Gradient method (see [2])
Permutations in W_{prior} are corrected with the spatio-temporal coherence method described in [2].

For the "domestic environment" datasets (i.e. the data used in the CHIME challenege), the batch ICA was applied to a signal chunk of 10s, while for the "public environment" the batch ICA was applied to the entire signal (i.e. still 10s).

The weights for the weighted Natural Gradient are determined according to the coherence of each frame in the TDOA direction of the target source (estimated through the Generalized State Coherence Transform [3] computed from each STFT input frame [2])

The on-line Semi-Blind Source Extraction algorithm in [1] processes the audio files frame-by-frame. The SBSS block in [1] uses filter lengths of 16384 and 4096 taps for the domestic and public environment data, respectively.  \mu was set to 0.2 while the step-size \eta was adapted from 0 to 0.2 according to the spatial coherence of each frame in the TDOA direction of the target. Set1 and Set2 indicates different settings for the step-size adaptation. The former applies a small variation to the step-size \eta, i.e. it is almost kept constant independently on the target source activity. The latter tends to reduce the step-size when the target source is active (where the activity is estimated from the coherence of the input data in the direction of W_prior).


[1]"Robust Automatic Speech Recognition through On-line Semi-Blind Source Extraction"
Francesco Nesta and Marco Matassoni, CHIME Workshop 2011, Florence(Italy)
http://spandh.dcs.shef.ac.uk/projects/chime/workshop/papers/pS22_nesta.pdf
[2]"Convolutive underdetermined source separation through weighted Interleaved ICA and spatio-temporal correlation", Francesco Nesta, Maurizio Omologo, submitted to LVA/ICA 2012, Tel Aviv 
[3] "Generalized State Coherence Transform for multidimensional TDOA estimation of multiple sources," Francesco Nesta, Maurizio Omologo, to appear on Audio, Speech, and Language Processing, IEEE Transactions on)