Contact and affiliations:
Jiri Malek (1), Zbynek Koldovsky (1) and Petr Tichavsky (2)

(1) Faculty of Mechatronic and Interdisciplinary Studies
Technical University of Liberec, Studentská 2, 461 17 Liberec, Czech
Republic

(2) Institute of Information Theory and Automation, Pod vodárenskou věží 4,
P.O. Box 18, 182 08 Praha 8, Czech Republic

jiri.malek(at)tul.cz

Description:

Our method is described in an article submitted to LVA/ICA 2012:
Semi-Blind Source Separation Based on ICA and Overlapped Speech Detection

Abstract:
We propose a semi-blind method for separation of stereo recordings of
two sources. The method assumes that a set of cancellation filters for
potential positions of one source (at least) is available. These filters
are computed from one-source-only intervals selected upon cross-talk
detection criteria. A stationary source in some of the assumed positions
is canceled by the corresponding filter, by which the other source is
separated. The stationary source is then separated by adaptive filter
that suppresses the formerly separated source. To select the appropriate
filter for the unknown position, we apply Independent Component Analysis
in a semi-blind fashion. The proposed method is verified on real-world
SiSEC data with stationary and moving sources.

Time of computation:
CPU frequency: 2.66 Ghz
Resulting times for Dataset 1:     Setup1:1017.94 s
                                                         Setup2:1000.66 s

i.e.

Resulting times per Ghz for Dataset 1:      Setup1: 2708 s

Setup2: 2661 s


Resulting times for Dataset 2:     Setup1:    36.47 s
                                                         Setup2: 190 s

i.e.

Resulting times per Ghz for Dataset 2:     Setup1: 97 s

Setup2: 505 s


There is a notable difference in Dataset 2 between Setup 1 a Setup 2.
This is not an error.
Our method computes cancellation filters for all fixed source positions
in the recordings and then selects the appropriate filter from the
available set is based on ICA.
In Setup 1 there is only single fixed source position, so only one
filter needs to be computed and it is applied automatically on the whole
recording without any blind ICA decision, which leads to significant
computational savings.