| dc.description.abstract | Symmetric adaptive decorrelation (SAD) is a semi-blind method of separating convolutely
mixed signals. While it has restrictions on the physical layout of the demixing equipment,
restrictions not present for many other blind source separation (BSS) techniques, it is more
ideally suited for some applications (for example, live sound mixing) due to the fact that
no post-processing is required to ascertain which output corresponds with which source.
Since the SAD algorithm is based on second-order statistics (SOS), it also tends to have a
lower computational load when compared with those based on higher order statistics. In
order to increase the e ciency of the SAD algorithm, a multibranched recursive structure is
investigated. While a nominal gain in e ciency is attained, we move away from this approach
in pursuit of more substantial gains. A frequency-domain symmetric adaptive decorrelation
(FD-SAD) algorithm is proposed, with savings increasing not only with larger lter sizes,
but also with increasing the number of sources. The convergence and stability of this new
algorithm is proven. A trade-o of the FD-SAD algorithm is that it introduces a delay in
the output, which renders the algorithm unsuitable for real-time applications. Therefore a
hybrid approach is also proposed that does not su er from the lag of the frequency domain
approach. While the proposed algorithm is slightly less computationally e cient than the
pure frequency domain algorithm, it is signi cantly more e cient than the time-domain
approach. A comparison of the frequency domain and hybrid algorithms shows that both
achieve separation equivalent to the time-domain algorithm in a real-world environment. The
proposed adaptations could also be used to extend other BSS approaches (such as Triple-N
ICA for Convolutive mixtures (TRINICON) [1], which can also be based on SOS), and a
comparison of the proposed methods with TRINICON is explored. | en_US |
