International Journal of Computational
Intelligence Research (IJCIR)
Volume 3, Number 1 (2007)
Multistage blind source separation and deconvolution for convolutive mixture
of speech signals
Yanxue Liang, Ichiro Hagiwara
Department of Mechanical Sciences and Engineering Graduate School of Science and Engineering of Tokyo Institute of Technology, Japan.
State Key Laboratory of Vibration, Shock & Noise, Shanghai Jiao Tong University
It is well known that conventional Multichannel Blind Deconvolution (MBD) suffers whitening effect. In present paper, MBD based multistage method is proposed to solve this problem. In detail, conventional MBD is first implemented, then compensation is conducted to estimate contributions of each source to every microphone, finally, a number of Single Input and Multi-Output (SIMO) dereverberation are carried out to recover original signal. On the other hand, time domain MBD algorithm is difficult to converge without good initialization. To deal with this problem, a new scheme in which FastICA based Direction of Arrival (DOA) estimation combines with Null Beamforming (NBF) is proposed. Such initialization generally guarantees convergence of time domain MBD so that compensation matrix can be constructed stably. Finally, experiment demonstrates validity and superiority of our scheme over other methods.
Blind Source Separation, Deconvolution, Multistage, FastICA, Null Beamforming, Spectral Compensation.