'Sparsification' of Audio Signals using the MDCT/lntMDCT and a Psychoacoustic Model - Application to Informed Audio Source Separation

摘要

Sparse representations have proved a very useful tool in a variety of domain, e.g. speech/music source separation. As strictly sparse representations (in the sense of e~0) are often impossible to achieve, other ways of studying signals sparsity have been proposed. In this paper, we revisit the irrelevance filtering analysis-synthesis approach proposed in (Balazs et al., IEEE Trans. ASLP, 18(1), 2010), where the TF coefficients that are below some masking threshold are set to zero. Instead of using the Gabor transform and a specific psychoacoustic model, we use tools directly inspired from perceptual audio coding, for instance MPEG-AAC. We show that significantly better "sparsification performances" are obtained on music signals, at lower computational cost. We then apply the sparsification process to the informed source separation (ISS) problem and show that it enables to significantly decrease the computational cost at the ISS decoder.