Blind Source Separation with Microphone Arrays
İlker Bayram, Istanbul Technical University, Istanbul, Turkey
Abstract
Blind dereverberation is a challenging problem due mainly to the nature of
the room impulse responses. A relatively recent approach for tackling this
problem is to use a microphone array and employ a multi-channel linear pre-
diction scheme. The idea is to linearly predict a reference observation using
delayed versions of all of the observations, with the residual forming the dere-
verbed estimate. In this talk, we consider a scenario where there are multiple
sources in the room, with unknown positions and room impulse responses.
We consider the application of the multi-channel linear prediction scheme for
such a scenario. We argue that, provided a certain assumption on the room
impulse responses is valid, such a preprocessing step transforms the observa-
tions so that they appear as if they were recorded in an anechoic environment.
This greatly simpli es the problem, turning the original reverberant source
separation problem into a non-reverberant one. In turn, simple beamforming
algorithms for source separation in anechoic environments can achieve sat-
isfactory performance. We demonstrate experimentally that the assumption
on the room impulse responses is valid, and provide experiments with real
recordings.
Speaker Biography
Ilker Bayram received the B.Sc. and M.Sc. degrees in Electrical and Elec-
tronics Engineering from Middle East Technical University (METU), Ankara,
Turkey, in 2002, and 2004 respectively. He recevied the Ph.D. degree in Elec-
trical Engineering from Polytechnic Institute of New York University in 2009.
In the following year, he was with the Biomedical Imaging Group at Ecole
Polytechnique Federale de Lausanne (EPFL), Switzerland, as a post-doctoral
researcher. In 2010, he joined Istanbul Technical University, Dept. of Electron-
ics and Communications Engineering, as an assistant professor. He is currently
an associate professor in the same department. His research interests are in
time-frequency frames, sparse signal processing, algorithms for reconstruction
problems, and machine learning for signal processing applications.