Multi-pitch estimation and tracking using Bayesian inference in block sparsity

In this paper, we consider the problem of multi-pitch estimation and tracking of an unknown number of harmonic audio sources. The regularized least-squares is a solution for simultaneous sparse source selection and parameter estimation. Exploiting block sparsity, the method allows for reliable tracking of the found sources, without posing detailed a priori assumptions of the number of harmonics for each source. The method incorporates a Bayesian prior and assigns data-dependent reg-ularization coefficients to efficiently incorporate both earlier and future data blocks in the tracking of estimates. In comparison with fix regularization coefficients, the simulation results, using both real and synthetic audio signals, confirm the performance of the proposed method.