Summary: A NEURAL NETWORK FOR SOUND SOURCE SEPARATION
JÖRN ANEMÜLLER AND TINO GRAMSS (
Graduate School in Psychoacoustics and Department of Physics
Carl von Ossietzky University, Oldenburg, Germany
The brain possesses the remarkable capability to filter incoming signals of multiple
speakers in such a way that the subject's attention can be focused on a single sound
source, the other sources being suppressed.
Much effort has been spent in order to mimic this behaviour by machines or
clever algorithms which try to reconstruct the separated sources. A key issue in such
attempts is the question `Which is the quantity to optimise', i.e., which function tells
us whether the original sources have been separated. An appealing answer to this
question might be given by the concept of Independent Component Analysis. Here,
the mutual statistical independence of the different source signals is exploited. I.e.,
one of the source signals does not give us information about any of the other sources.
However, when different mixtures of the source signals are detected, e.g., at the
two ears of a human subject, the mixtures are strongly correlated, i.e., not indepen-