Blind Source Separation for Convolutive Mixtures with Neural Networks

Abstract

Blind source separation of convolutive mixtures is used as a preprocessing stage in many applications. The aim is to extract individual signals from their mixtures. In enclosed spaces, due to reverberation, audio signal mixtures are considered to be convolutive ones. Time domain algorithms (as neural network based blind source separation) are not suitable for signal recovery from convolutive mixtures, thus the need of frequency domain or subband processing arise. We propose a subband approach: first the mixtures are split to several subbands, next time-domain blind source separation is carried out in each subband, finally the recovered sources are recomposed from the subbands. The major drawback of the subband approach is the unknown order of the recovered sources. Regardless of this undesired phenomenon the subband approach is faster and more stable than the simple time domain algorithm.

Keywords

• blind source separation
• neural networks
• independent component analysis
• subband analysis and synthesis