Audio source separation with convolutively mixed signals

Abstract

In this thesis, we addressed the problem of audio source separation of convolutivelymixed signals using microphone arrays. Independent Component Analysisis a major statistical tool for solving this problem. In real room environments, the recordings of audio signals usually involve the signal itself as well as some delayed andamplitude modulated versions of this signal. This is due to reverberation or echo ofthe room which occurs as a result of reflection of walls, ceiling, ground as well as thefurniture inside. Separation of signals that are mixed in these kinds of environments is a challenging problem. There exist both time-domain and frequency-domain solutionsto this problem. We mostly focus on frequency domain methods where ICA isperformed separately in each frequency bin. Permutation ambiguity which is the basicproblem in frequency domain ICA, is also handled with two basic approaches which are, direction of arrival method which is motivated by conventional beamforming theoryand interfrequency correlations which is motivated by nonstationarity of speechsignals. Conventional and Adaptive Beamforming methods are also implemented here.These methods separate sources by exploiting the physics of the propagation.