Subband-based Single-channel Source Separation of Instantaneous Audio Mixtures
Jalil Taghia, Mohammad Ali Doostari · 2009
Abstract: In this paper, a new algorithm is developed to separate the audio sources from a single instantaneous mixture. The algorithm is based on subband decomposition and uses a hybrid system of Empirical Mode Decomposition (EMD) and Principle Component Analysis (PCA) to construct artificial observations from the single mixture. In the separation stage of algorithm, we use Independent Component Analysis (ICA) to find independent components. At first the observed mixture is divided into a finite number of subbands through filtering with a parallel bank of FIR band-pass filters. Then EMD is employed to extract Intrinsic Mode Functions (IMFs) in each subband. By applying PCA to the extracted components, we find uncorrelated components which are the artificial observations. Then we obtain independent components by applying Independent Component Analysis (ICA) to the uncorrelated components. Finally, we carry out subband synthesis process to reconstruct fullband separated signals. The experimental results substantiate that the proposed method truly performs the task of source separation from a single instantaneous mixture. Key words: Blind source separation • independent component analysis • empirical mode decomposition • principle component analysis • single-channel audio source separation • subband decomposition