Performance of Permutation-Free ICA

LEANDRO DI PERSIA; TADASHIGE NOGUCHI; KENKO OHTA; MASUZO YANAGIDA

Kyoto

Encuentro; Monthly meeting of the Special Interest Group on Speech; 2006

Institute of Electronics, Information and Communication Engineering

Performance evaluation of Permutation-Free ICA for blind source separation is presented. Permutation-Free ICA was proposed as a method that couldavoid the so-called "Permutation problem" that appears in frequency-domain ICA from the necessity to align components of each source. Up to now, several approaches for the permutation problem have been proposed.ç, for example those based on directivity patterns of each frequency bin, those on correlation between adjacent frequency bins and those based on harmonic structure of source sounds. These methods, however, do not give the perfect solution to the permutation problem. On the other hand, Permutation-Free ICA carries out separation on a set of long vectors consisting of connected temporal changes of frequency components of the received signal. Investigating the performance of Permutation-Free ICA, it appears that the method is effective for cases where short-time speech or nonstationary noise are involved. The separation matrix obtained in the Permutation-Free ICA, however, has a commont directivity pattern for all frequency. To make Permutation-Free ICA effective for real situations, some means that can find the optimum separation matrix for each frequency bin referring to the results of Permutation-Free ICA.