Convolutive Blind Source Separation for Noisy Mixtures

Convolutive blind source separation (BSS) is a promising technique for separating acoustic mixtures acquired by multiple microphones in reverberant environments. In contrast to conventional beamforming methods no a-priori knowledge about the source positions or sensor arrangement is necessary resulting in a greater versatility of the algorithms. In this contribution we will first review a general BSS framework called TRINICON which allows a unified treatment of broadband and narrowband BSS algorithms. Efficient algorithms will be presented and their high performance will be confirmed by experimental results in reverberant rooms. Subsequently, the BSS model will be extended by incorporating background noise. Commonly encountered realistic noise types are examined and, based on the resulting model, pre-processing methods for noise-robust BSS adaptation are investigated. Additionally, an efficient post-processing technique following the BSS stage, will be presented, which aims at simultaneous suppression of background noise and residual cross-talk. Combining these pre- or post-processing approaches with the algorithms obtained by the TRINICON framework yield versatile BSS systems which can be applied in adverse environments as will be demonstrated by experimental results.