Semi-Blind Audio Source Separation of Linearly Mixed Two-Channel Recordings via Guided Matching Pursuit
This paper describes a source separation system with the intent to be used in high quality audio post-processing tasks. The system is to be used as the front-end of a larger system capable of modifying the individual sources of existing, two-channel, multi-source recordings. Possible applications include spatial re-configuration such as up-mixing and pan-transformation, re-mixing, source suppression/elimination, source extraction, elaborate filtering, timestretching and pitch-shifting. The system is based on a new implementation of the Matching Pursuit algorithm and uses a known mixing matrix. We compare the results of the proposed system with those from mpd-demix of the ’MPTK’ software package and show that we get similar evaluation scores and in some cases better perceptual scores. We also compare against a segmentation algorithm which is based on the same principles but uses the STFT as the front-end and show that source separation algorithms based on adaptive decomposition schemes tend to give better results. The novelty of this work is a new implementation of the original Matching Pursuit algorithm which adds a pre-processing step into the main sequence of the basic algorithm. The purpose of this step is to perform an analysis on the signal and based on important extracted features (e.g frequency components) create a mini-dictionary comprising atoms that match well with a specific part of the signal, thus leading to focused and more efficient exhaustive searches around centres of energy in the signal.