Guided Matching Pursuit and its Application to Sound Source Separation

In the last couple of decades there has been an increasing interest in the application of source separation technologies to musical signal processing. Given a signal that consists of a mixture of musical sources, source separation aims at extracting and/or isolating the signals that correspond to the original sources. A system capable of high quality source separation could be an invaluable tool for the sound engineer as well as the end user. Applications of source separation include, but are not limited to, remixing, up-mixing, spatial re-configuration, individual source modification such as filtering, pitch detection/correction and time stretching, music transcription, voice recognition and source-specific audio coding to name a few. Of particular interest is the problem of separating sources from a mixture comprising two channels (2.0 format) since this is still the most commonly used format in the music industry and most domestic listening environments. When the number of sources is greater than the number of mixtures (which is usually the case with stereophonic recordings) then the problem of source separation becomes under-determined and traditional source separation techniques, such as “Independent Component Analysis” (ICA) cannot be successfully applied. In such cases a family of techniques known as “Sparse Component Analysis” (SCA) are better suited. In short a mixture signal is decomposed into a new domain were the individual sources are sparsely represented which implies that their corresponding coefficients will have disjoint (or almost) disjoint supports. Taking advantage of this property along with the spatial information within the mixture and other prior information that could be available, it is possible to identify the sources in the new domain and separate them by going back to the time domain. It is a fact that sparse representations lead to higher quality separation. Regardless, the most commonly used front-end for a SCA system is the ubiquitous short-time Fourier transform (STFT) which although is a sparsifying transform it is not the best choice for this job. A better alternative is the matching pursuit (MP) decomposition. MP is an iterative algorithm that decomposes a signal into a set of elementary waveforms called atoms chosen from an over-complete dictionary in such a way so that they represent the inherent signal structures. A crucial part of MP is the creation of the dictionary which directly affects the results of the decomposition and subsequently the quality of source separation. Selecting an appropriate dictionary could prove a difficult task and an adaptive approach would be appropriate. This work proposes a new MP variant termed guided matching pursuit (GMP) which adds a new pre-processing step into the main sequence of the MP algorithm. The purpose of this step is to perform an analysis of the signal and extract important features, termed guide maps, that are used to create dynamic mini-dictionaries comprising atoms which are expected to correlate well with the underlying signal structures thus leading to focused and more efficient searches around particular supports of the signal. This algorithm is accompanied by a modular and highly flexible MATLAB implementation which is suited to the processing of long duration audio signals. Finally the new algorithm is applied to the source separation of two-channel linear instantaneous mixtures and preliminary testing demonstrates that the performance of GMP is on par with the performance of state of the art systems