Blind Detection of CLT Disobeying Frequency Bind for Audio Source Separation by Fixed-Point ICA


NAIST \| Library	Advanced Search	Japanese \| English

Please use this identifier to cite or link to this item: http://hdl.handle.net/10061/8019

Title:	Blind Detection of CLT Disobeying Frequency Bind for Audio Source Separation by Fixed-Point ICA
Authors:	Raj Kishore Prasad Hiroshi Saruwatari Kiyohiro Shikano
Issue Date:	Apr-2004
Start page:	IV3151
End page:	IV3154
Abstract:	This paper peeps into the problem of disobeyance of Central Limit Theorem (CLT) by the Time-Frequency Series of the Speech (TFSS) data. For the blind separation algorithms, based on the non-gaussianization of mixed signal e.g. signal separation by negentropy or kurtosis maximization, such behavior by TFSS is problematic in the sense that it is against the working spirit of algorithm. The detection of such CLT failure bin is important. This paper describes (explores) such failure in depth and prescribes a blind method, based on the statistics TFSS that can detect 70-75% of the CLT disobeying frequency bins. The proposed method is blind in the sense that it relies only on the statistics of observed data.
Description:	ICA2004: 18th International Congress on Acoustics, April 4-9, 2004, Kyoto, Japan.
URI:	http://hdl.handle.net/10061/8019
Text Version:	Publisher
Appears in Collections:	情報科学研究科 / Graduate School of Information Science

Files in This Item:

File	Size	Format
ICA_2004_IV3151.pdf	602.28 kB	Adobe PDF	View/Open