Extraction of independent non-Gaussian components from non-Gaussian background
Zbyněk Koldovský · The Journal of the Acoustical Society of America · 2018
Independent component analysis (ICA) is a popular method for Blind Signal Separation applied to multichannel linear recordings. Convolutive mixtures of acoustic signals can be separated by applying ICA in the frequency domain to each sub-band separately or jointly as in Independent Vector Analysis (IVA). However, this way, the mixtures are separated into that many components as is the number of sensors. ICA and IVA are therefore not effective in applications where only one (or few) signals are of practical interest. We introduce a new parameterization of the ICA mixing model that is optimized for the extraction of one signal of interest (SOI). The approach is called Independent Component Extraction (ICE) and is closely related to methods for Blind Signal Extraction (BSE) such as One-unit FastICA. However, BSE methods assume that background signals (the other signals than SOI) are all Gaussian, which limits their accuracy. In this work, we show that ICE can be extended also for a non-Gaussian background so that the accuracy of algorithms is improved, although all components are not separated as in ICA. We also introduce an extension for the extraction of a vector component, so-called Independent Vector Extraction.