Blind Source Separation and Speech Enhancement

Abstract

Humans are able to pay attention to a particular speaker in a complex auditory scene of multiple talkers. This amazing functionality has motivated the development of the blind source separation (BSS) methods that aim to recover source signals from their mixtures without detailed knowledge of the mixing process. A problem of this kind is also known as the cocktail party problem. We shall consider two scenarios, one is to recover all possible sources, and the other is to enhance one signal while suppressing the rest. Because the mixing of sound signals are nonlocal (convolutive) in time, we shall first consider the simpler instantaneous mixture problems where solutions are more explicit for better illustration of general ideas.