Doctoral Dissertation High-Fidelity Blind Source Separation Using Single-Input-Multiple-Output-Model-Based Independent Component Analysis

Blind source separation (BSS) technique using independent component analysis (ICA) for acoustic signals has been developed over the last decade. This technique assumes that the source signals are mutually independent, and can estimate the source signals from the mixed signals without a priori information. Thus, this technique is highly applicable in high-quality hands-free telecommunication system. The conventional ICA-based BSS method is a means of extracting the independent sound source signals as the monaural signals from the mixed signals observed in each input channel, and the separated signals include arbitrary spectral distortions. Consequently, they have a serious drawback in that the separated sounds cannot maintain information about the directivity, localization, reverberation, or spatial qualities of each sound source. These problems prevent any BSS methods from being applied to binaural signal processing or high-fidelity sound reproduction system. In this thesis, firstly, in order to solve the above-mentioned fundamental problems, we propose a new ICA algorithm, which is called Single-Input MultipleOutput (SIMO)-model-based ICA (SIMO-ICA) with least squares criterion (SIMOICA-LS). Here the term ”SIMO” denotes specific transmission system in which the input is a single source signal and the outputs are its transmitted signals observed at multiple sensors. SIMO-ICA-LS can separate the mixed signals, not into ∗ Doctoral Dissertation, Department of Information Processing, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DD0361208, March 24, 2006.