Harmonic-Percussive Source Separation Using Harmonicity and Sparsity Constraints

In this paper, we propose a novel approach to harmonicpercussive sound separation (HPSS) using Non-negative Matrix Factorization (NMF) with sparsity and harmonicity constraints. Conventional HPSS methods have focused on temporal continuity of harmonic components and spectral continuity of percussive components. However, it may not be appropriate to use them to separate time-varying harmonic signals such as vocals, vibratos, and glissandos, as they lack in temporal continuity. Based on the observation that the spectral distributions of harmonic and percussive signals differ – i.e., harmonic components have harmonic and sparse structure while percussive components are broadband – we propose an algorithm that successfully separates the rapidly time-varying harmonic signals from the percussive ones by imposing different constraints on the two groups of spectral bases. Experiments with real recordings as well as synthesized sounds show that the proposed method outperforms the conventional methods.