Music Source Separation via Sparsified Dictionaries vs. Parametric Models

In the framework of audio signal analysis, there have been recent significant advances in two directions: sparse and structured representation. In fact, sparse decompositions of audio signals are shown to be effective, and appear to be extremely useful in many signal processing applications: compression, source separation, noise reduction... A second point of view tries to take advantage of the harmonic structure of the audio signal. It models a note signal as a periodic signal with (slow) global variation of amplitude (reflecting attack, sustain, decay) and frequency (limited time warping). In this paper, we compare the two approaches through experiments involving various audio signals. We consider particularly application of the two approaches to noise reduction, and underdetermined Blind Source Separation.