- Andreas Schwarz
The problem of recovering the original signals and determining the source positions from an observed linear mixture can be addressed using Blind Source Separation (BSS) techniques. The term “blind” implies that the mixing process and the original signals are unknown. Based on blind adaptive Multiple-Input-Multiple-Output (MIMO) system identification, and originally developed as an adaptive BSS algorithm, an Acoustic Source Localization (ASL) scheme has been developed under Matlab and implemented as a real-time demonstrator written in C/C++ language.
The algorithm allows the localization and separation of several simultaneously active sound sources, possibly in several dimensions. Traditionally, BSS algorithmic performance is assessed in terms of Signal-to-Interference Ratio (SIR). When measured successively at the microphone inputs and at the BSS outputs, the gain in SIR actually measures the degree of source separation obtained by applying the adaptive BSS demixing filters. The SIR offers therefore valuable knowledge on the convergence state of the BSS algorithm, which could be advantageously used, e.g., to provide an efficient step-size control mechanism. Shadow filtering approaches constitute another important domain of application of such a knowledge, where the performance of several separating entities (typically, a set of BSS algorithms running in parallel) need to be compared to be able to blindly select the best candidate.
However, the SIR cannot be measured in real applications since it necessitates some a-priori knowledge on the acoustical mixing process. In this thesis, alternative “blind” performance measures built upon available signals should be investigated. An experimental study conducted under Matlab will show how these measures reflect the separation and localization performance of the algorithm, and promising control mechanisms based on such measures should be identified.