BLIND SEPARATION OF DISJOINT ORTHOGONAL SIGNALS:
DEMIXING N SOURCES FROM 2 MIXTURES
Alexander Jourjine
Siemens Corporate Research
755 College Road East
Princeton, NJ 08540, USA
f
jourjine,rickard
g
@scr.siemens.com
Scott Rickard
Ozgur Ylmaz
Princeton University
Fine Hall, Washington Road
Princeton, NJ 08540, USA
f
srickard,oyilmaz
g
@princeton.edu
ABSTRACT
We present a novel method for blind separation of any num-
ber of sources using only two mixtures. The metho d applies
when sources are (W-)disjoint orthogonal, that is, when the
supports of the (windowed) Fourier transform of any two
signals in the mixture are disjoint sets.
We show that, for anechoic mixtures of attenuated and
delayed sources, the metho d allows one to estimate the mix-
ing parameters by clustering ratios of the time-frequency
representations of the mixtures. The estimates of the mix-
ing parameters are then used to partition the time-frequency
representation of one mixture to recover the original sources.
The technique is valid even in the case when the number
of sources is larger than the number of mixtures. The gen-
eral results are veried on both speech and wireless signals.
Sample sound les can be found here:
http://www.princeton.edu/~srickard/bss.html
1. INTRODUCTION
Demixing noisy mixtures has been a goal of long stand-
ing in the eld of blind source separation(BSS). One area
where BSS methods are imp ortant is wireless communica-
tions where receiving antennas measure a linear mixture of
delayed and attenuated EM radiation of the source signals.
Another example lies in the acoustic domain where it is de-
sirable to separate a voice of interest from background noise
and interfering speakers.
Assumptions on the statistical properties of the sources
usually provide a basis for a demixing algorithm. Some
common assumptions are that the sources are statistically
independent[1 , 2], are statistically orthogonal[3 ], are non-
stationary[4 ], or can be generated by nite dimensional
model spaces[5 ]. The independence/orthogon ality assump-
tion can be veried exp erimentally for speech signals and is
also valid for many wireless communications schemes. Some
of these methods work well for instantaneous demixing, but
fail if propagation delays are present. Additionally, many
algorithms are computationally intensive as they require the
computation of higher-order statistical moments or the op-
timization of a non-linear cost function.
One area of research in blind source separation that is
relatively untouched is when there are less mixtures than
sources. We refer to such a case as
degenerate blind source
separation
. Degenerate blind source separation poses a chal-
lenge because the mixing matrix is not invertible and the
traditional metho d of demixing by estimating the inverse
mixing matrix does not work. As a result, most of re-
search in channel estimation and BSS has b een done for the
square non-degenerate case. In the related areas of wireless
communication where channel estimation is important, the
number of receivers is typically more then the number of
emitters. For example, subspace channel estimation meth-
ods require at least one more mixture than sources to esti-
mate the sources and the noise[6, 7].
One example of degenerate blind source separation esti-
mates an arbitrary number of sources from a single mixture
by modeling the signals as AR pro cesses[8 ]. However, this is
achieved at a price of approximating signals by AR sto chas-
tic processes, which can be to o restrictive. Another exam-
ple of degenerate separation uses higher order statistics to
demix three sources from two mixtures[9 ]. This approach
is not feasible however for a large number of sources since
the use of higher order statistics of mixtures leads to an
explosion in computational complexity.
Similar in spirit to this paper, van Hulle employs a clus-
tering method for relative amplitude parameter estimation
and degenerate demixing[10 ]. The assumptions used by van
Hulle were that only one signal at a given time is non-zero
and that mixing is instantaneous, that is, there is only a
relative amplitude mixing parameter associated with each
source. In real world acoustic environments or wireless com-
munication domains, these assumptions are not valid.
The results of this paper are derived for anechoic time
delay mixtures. We prove that for such a mixing mo del,
estimation of the mixing parameters and complete separa-
tion of any number of disjoint orthogonal sources is p ossible
from only two mixtures. The results can b e extended to the
noisy echoic case[11].
In Section 2 we dene the time delay mixing mo del, in-
troduce the concept of disjoint orthogonality, and describ e
the mixing parameter estimation. In Section 3 we describ e
a solution for degenerate demixing.
2. MIXING PARAMETER ESTIMATION
2.1. Source mixing
Consider measurements of a pair of sensors where only the
direct path is present. In this case, without loss of general-
To app ear in
Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP2000).
1