Cc. Ko et Hk. Garg, UNBIASED SEPARATION AND ENHANCEMENT OF SOURCES BY USING A ZERO TRACKING ALGORITHM WITH POST-PROCESSING ADAPTIVE NOISE-REDUCTION, Digital signal processing, 7(4), 1997, pp. 209-221
This paper proposes and investigates a new structure for the unbiased
separation, tracking, and enhancement of uncorrelated sources by a lin
ear array. The structure consists of a matrix preprocessing beamformer
followed by an adaptive linear combiner and a postprocessing beam-for
mer. The weights of the adaptive linear combiner are updated by using
the LMS algorithm and, together with the preprocessor, implement an ad
aptive array whose response zeros are adjusted to minimize the output
power with unity element weight vector norm constraint. By properly de
signing the preprocessor, first-order movements in the response zeros
can be individually and proportionally controlled by first order chang
es in the weights in the linear combiner, and the preprocessor outputs
will be due to individual sources and receiver noise in the steady st
ate. By measuring the powers of the preprocessor outputs, it will be p
ossible to determine if a particular preprocessor output is due to an
actual source (if the associated zero is tracking a source) or merely
receiver noise (if the associated zero is not tracking any source). Fu
rther SNR enhancement is then carried out by making use of those prepr
ocessor outputs due to receiver noise to remove the correlated receive
r noise components in the other preprocessor outputs tracking actual s
ources through the use of a matrix postprocessor. Apart from the imple
mentation of the preprocessor and postprocessor which will be inevitab
le in any source separation system and which have to be designed only
occasionally and whenever the response zeros have been found to have c
hanged significantly, the algorithm has an implementation complexity w
hich is proportional to the array size. However, as the preprocessor o
utputs are due primarily to individual sources in the steady state, th
e algorithm has only one single asymptotic time constant controlled by
the designed misadjustment level and has fast tracking performance. (
C) 1997 Academic Press.