Single channel blind source separation

Abstract

Single channel blind source separation (SCBSS) is an intensively researched field with numerous important applications. This research sets out to investigate the separation of monaural mixed audio recordings without relying on training knowledge. This research proposes a novel method based on variable regularised sparse nonnegative matrix factorization which decomposes an information-bearing matrix into two-dimensional convolution of factor matrices that represent the spectral basis and temporal code of the sources. In this work, a variational Bayesian approach has been developed for computing the sparsity parameters of the matrix factorization. To further improve the previous work, this research proposes a new method based on decomposing the mixture into a series of oscillatory components termed as the intrinsic mode functions (IMF). It is shown that IMFs have several desirable properties unique to SCBSS problem and how these properties can be advantaged to relax the constraints posed by the problem. In addition, this research develops a novel method for feature extraction using psycho-acoustic model. The monaural mixed signal is transformed to a cochleagram using the gammatone filterbank, whose bandwidths increase incrementally as the center frequency increases; thus resulting to non-uniform time-frequency (TF) resolution in the analysis of audio signal. Within this domain, a family of Itakura-Saito (IS) divergence based novel two-dimensional matrix factorization has been developed. The proposed matrix factorizations have the property of scale invariant which enables lower energy components in the cochleagram to be treated with equal importance as the high energy ones. Results show that all the developed algorithms presented in this thesis have outperformed conventional methods.