A method for solving the permutation problem in ICA

LEANDRO DI PERSIA; KENKO OHTA; MASUZO YANAGIDA

Tokio

Congreso; Monthly meeting of the Special Interest Group on Spoken Language Processing; 2006

Institute of Electronics, Information and Comunication Engineering, Japan

Proposed is a method for solving the permutation problem in frequency-domain independent component analysis (ICA). Up to now, several approaches for the permutation problem are proposed, for example those based on directivity pattern of each frequency bin, the correlation between frequency bins and harmonic structure. These methods, however, don not give the complete solution. So, it is necessary to develop a method which solves or does not need to solve the permutation problem. The traditional frequency-domain ICA executes separation using a matrix consisting of the short time Fourier transform of the received signal. The proposed method, however, carries out separation on a long vector constructed as connected frequency bins of the received signal. Though, there is a possibility that whole the channel of the separated signal might be exchanged with the other channel, there is no possibility that a single or a group of channels are permuted with other channels.