Abstract

Acoustic source separation is a relatively recent topic of signal processing which aims to simultaneously separate many acoustic sources recorded through one or more microphones. Such a problem was formulated to emulate the natural capability of the human auditory system which is able to recognize and enhance the sound coming from a particular source. Addressing this problem is of high interest in the automatic speech recognition (ASR) community since it would improve the effectiveness of a natural human-machine interaction. Among numerous methods of multichannel blind source separation techniques, those based on the Independent Component Analysis (ICA) applied in the frequency-domain [81] are the most investigated, due to their straightforward physical interpretation and computational efficiency. In spite of recent developments many issues still need to be address to make such techniques robust in adverse conditions, such as high reverberation, ill-conditioning and occurrence of permutations. Furthermore, most of the proposed BSS methods are computationally expensive and not feasible for a real-time implementation. This PhD thesis describes a research activity in the robust separation of acoustic sources in adverse environment. A new framework of blind and semi-blind techniques is proposed which allows source localization and separation even in highly reverberant environment and with realtime constraint. For each proposed technique, theoretical and practical issues are discussed and a comparison with alternative state-of-art methods is provided. Furthermore, the robustness of the proposed framework is validated implementing two real-time blind and semi-blind systems which are tested in challenging real-world scenarios.