Permutation Correction and Speech Extraction Based on Split Spectrum Through FastICA
Hiromu Gotanda, Kazuyuki Nobu, Takeshi Koya, Kei-ichi Kaneda, Toshihiro Ishibashi, Naomi Haratani · 2003
A blind source deconvolution method without indeterminacy of permutation and scaling is proposed by using notable features of split spectrum and locational information on signal sources. A method for extracting human speech exclusively is also proposed by taking advantage of the rule, the property of FastICA separates sources in order of large non-Gaussianity from their mixtures and the fact that human speeches are usually larger in non-Gaussianity than noises. The proposed methods have been veried by several experiments in a real room.