This is an approach to the modelling of time-frequency surfaces which consists of a Bayesian regularization scheme in which the prior distributions over the time-frequency coefficients are constructed to favour smoothness of the estimated function and sparseness of the coefficient representation both.