Accurate, fast and stable denoising source
separation algorithms
Harri Valpola1,2? and Jaakko Sarela2??
1 Artificial Intelligence Laboratory, University of Zurich
Andreasstrasse 15, 8050 Zurich, Switzerland
2 Neural Networks Research Centre, Helsinki University of Technology
P.O.Box 5400, FI-02015 HUT, Espoo, Finland
[email protected]
Abstract. Denoising source separation is a recently introduced frame-
work for building source separation algorithms around denoising pro-
cedures. Two developments are reported here. First, a new scheme for
accelerating and stabilising convergence by controlling step sizes is in-
troduced. Second, a novel signal-variance based denoising function is
proposed. Estimates of variances of different source are whitened which
actively promotes separation of sources. Experiments with artificial data
and real magnetoencephalograms demonstrate that the developed algo-
rithms are accurate, fast and stable.
1 Introduction
In denoising source separation (DSS) framework [1], separation algorithms are
built around a denoising function. This makes it easy to tailor source separation
algorithms for the task at hand. Good denoisings usually result in fast and
accurate algorithms. Furthermore, explicit ob jective function is not needed, in
contrast to most existing source separation algorithms.
Here we report further developments of two aspects. First, we introduce a new
method for stabilising and accelerating convergence which is inspired by predic-
tive controllers. Second, we develop further the signal-variance-based denoising
principles. The resulting algorithms yield good results in terms of signal-to-noise
ratio (SNR) and exhibit fast and stable convergence.
2 Source separation by denoising
Consider a linear instantaneous mixing of sources:
X=AS+ν,
(1)
? Funded by the European Commission, under the project ADAPT (IST-2001-37173)
and by the Academy of Finland, under the project New information processing
principles.
?? Funded by the Academy of Finland.