ON-LINE TIME-DOMAIN BLIND SOURCE SEPARATION OF NONSTATIONARY CONVOLVED SIGNALS

In this paper we propose a time-domain gradient algorithm that exploits the nonstationarity of observed signals and recovers the original sources by simultaneously decorrelating time-varying secondorder statistics. By introducing a generalized weighting factor in our cost function we can formulate an on-line algorithm that can be applied to time-varying multipath mixing systems. A further benefit is the possibility of implementing updates in a recursive manner and thus reduce computational complexity. We show that this method inherently possesses an adaptive step size and hence avoids stability problems. Furthermore we present a new geometric initialization for time-domain gradient algorithms that improves separation performance in strongly reverberant environments. In our experiments we compared the separation performance of the proposed algorithm with those of its off-line counterpart and another multiple-decorrelation-based on-line algorithm in the frequency domain for real-world speech mixtures.

[1]  S. Haykin.  Unsupervised adaptive filtering, vol. 1: Blind source separation , 2000 .

[2]  Kiyotoshi Matsuoka,et al.  A neural net for blind separation of nonstationary signals , 1995, Neural Networks.

[3]  Shoko Araki,et al.  Time domain blind source separation of non-stationary convolved signals by utilizing geometric beamforming , 2002, Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing.

[4]  Tetsunori Kobayashi,et al.  ASJ continuous speech corpus for research , 1992 .

[5]  Shun-ichi Amari,et al.  Natural Gradient Works Efficiently in Learning , 1998, Neural Computation.

[6]  Noboru Ohnishi,et al.  A method of blind separation for convolved non-stationary signals , 1998, Neurocomputing.

[7]  E. Oja,et al.  Independent Component Analysis , 2013 .

[8]  S. Haykin,et al.  Adaptive Filter Theory , 1986 .

[9]  Lucas C. Parra,et al.  On-line Convolutive Blind Source Separation of Non-Stationary Signals , 2000, J. VLSI Signal Process..

[10]  Shoko Araki,et al.  Equivalence between frequency domain blind source separation and frequency domain adaptive null beamformers , 2001, INTERSPEECH.

[11]  Jean-Francois Cardoso,et al.  Blind signal separation: statistical principles , 1998, Proc. IEEE.

[12]  Shiro Ikeda,et al.  A METHOD OF ICA IN TIME-FREQUENCY DOMAIN , 2003 .

[13]  M. Bellanger.  Adaptive filter theory: by Simon Haykin, McMaster University, Hamilton, Ontario L8S 4LB, Canada, in: Prentice-Hall Information and System Sciences Series, published by Prentice-Hall, Englewood Cliffs, NJ 07632, U.S.A., 1986, xvii+590 pp., ISBN 0-13-004052-5 025 , 1987 .

[14]  Christopher V. Alvino,et al.  Geometric source separation: merging convolutive source separation with geometric beamforming , 2001, Neural Networks for Signal Processing XI: Proceedings of the 2001 IEEE Signal Processing Society Workshop (IEEE Cat. No.01TH8584).

[15]  Kiyohiro Shikano,et al.  Comparison of time-domain ICA, frequency-domain ICA and multistage ICA for blind source separation , 2002, 2002 11th European Signal Processing Conference.