Sparse Bayesian representation in time–frequency domain

We consider a Bayesian time–frequency surfaces modeling of sound signals. The model is based on decomposing a signal into time–frequency domain using Gabor frames, which requires a careful regularization through appropriate variable selection to cope with the overcompleteness. We propose to impose a time-line beta-Bernoulli prior on the time–frequency coefficients of Gabor frames to create dependency structures coupled with the stochastic search variable selection to achieve sparsity. Theoretical aspects of the prior specification are investigated and an efficient MCMC algorithm is developed. Performance of the proposed model with other popularly used models is compared through analyzing simulated and real signals.