Analysis of the performance and limitations of ICA-based relative impulse response identification

Estimating impulse responses for a single source is a crucial problem for many applications in audio signal processing, such as source extraction. Since absolute impulse responses are hard to identify, relative impulse responses or, equivalently, relative transfer functions are identified instead. Independent Component Analysis (ICA) for convolutivemixtures offers the possibility to determine relative impulse responses implicitly by separating the target source from interfering sources. In this paper, fundamental limitations of relative transfer function (RTF) estimation are analyzed by calculating least-squares (LS)-optimal estimates in adverse scenarios, where the influence of scatterers and reverberation on the performance must be accounted for. Hereupon, ICA-based RTF estimation in the TRINICON framework is compared with the LS-optimal estimates.