Doctoral Dissertation Blind Source Separation Based on Multistage Independent Component Analysis

A hands-free speech recognition system and a hands-free telecommunication system are essential for realizing an intuitive, unconstrained, and stress-free human-machine interface. In real acoustic environments, however, the speech recognition performance and a speech recording performance significantly degraded because we cannot detect the user’s speech with a high signal-to-noise ratio (SNR) owing to the interference signals such as noise. In this thesis, we introduce blind source separation (BSS), which is an approach for estimating original source signals only from the information of the mixed signals observed in each input channel. Many BSS methods based on independent component analysis (ICA) have been proposed for the acoustic signal separation. However, the performances of these methods degrade seriously particularly under extreme reverberant conditions. The ICA-based BSS can be classified into two groups in terms of the processing domain, i.e., frequency-domain ICA (FDICA) and time-domain ICA (TDICA) From the experimental study using the conventional FDICA, the source-separation performance is saturated because the independence assumption collapses in each narrow-band. In TDICA, the convergence degrades because the iterative learning rule becomes more complicated as the reverberation increases. In order to resolve the problems, I newly propose multistage ICA (MSICA), in which FDICA and TDICA are cascaded. In the proposed method, the separated signals of FDICA ∗ Doctoral Dissertation, Department of Information Processing, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DD0261017, March 24, 2005.