| |
Abstract:
A parametric time-delay model of mixing is introduced. For N
sources it is defined by an NxN matrix of source attenuation
coefficients and an NxN matrix of delays in signal propagation
times. It is shown, using TITO problem as a basis, that N
statistically orthogonal sources can be separated blindly from N
time-delay mixtures using only second order statistics. The
separation is performed by estimation of the attenuation and delay
parameters, which is done by solving non-linear cross-correlation
constraints resulting from the statistical orthogonality of the
sources. The parameters are then used to construct a demixing
matrix of a known parametric form. A solution to the channel
selection problem is presented together with a method of removal of
demixing artifacts by application of Wiener filter. The results are
verified in an example of speech signals by using both time and
frequency domain algorithms.
|
