Francesco Nesta
Fondazione Bruno Kessler (FBK-IRST), Trento (Italy)

Main parameters:

FFT size=2048
Frame shift=256
Number of ICA iterations per frequency=20
Step-size=0.1

Task1
Sources have been separated with an algorithm based on a Regularized Recursive ICA (RR-ICA)[1] which extends the method in [2]. 
A not-well optimized matlab version of the algorithm takes about 15-20 seconds per mixture on a Intel Core Duo 2.1 Ghz.


Task2
The demixing parameters are estimated by the same algorithm of task1. The source separation is improved by performing a further time-varying spatial filtering on the estimated outputs (similar to that proposed in [3]).


[1] Francesco Nesta, Piergiorgio Svaizer, Maurizio Omologo, "Convolutive BSS of short mixtures by ICA recursively regularized across frequencies", (accepted) in IEEE  Transactions on Audio, Speech and Language Processing.
[2] Francesco Nesta, Piergiorgio Svaizer, Maurizio Omologo, "A BSS method for short utterances by a recursive solution to the permutation problem", (in proceedings), SAM2008
[3] H. Sawada, S. Araki, R. Mukai, S. Makino, "Blind Extraction of Dominant Target Sources Using ICA and Time-Frequency Masking," IEEE Trans. Audio, Speech, and Language Processing, vol.14, no.6, pp.2165-2173, Nov. 2006