Adaptive time-frequency representation for sound analysis and processing *

International audienceThe readability of a time-frequency representation generally depend crucially on a priori choices of some analysis parameters, which are often chosen heuristically. We describe here a set of methods for automatically selecting optimal parameters, exploiting sparsity requirements.Using Shannon or Renyi entropies for defining sparsity, we show that the “optimal” representation unfortunately often depends on the criterion. Therefore, we present several ways out for correcting such a shortcoming. In particular, we exploit the idea of “local optimization” in the time-frequency plane. Two different approaches are presented. In a supervised approach, the user can manually select the time-frequency domain in which window optimization is to be performed. In the unsupervised case, an iterative algorithm yields an automatic time-frequency segmentation, together with a corresponding inversion algorithm. As a by-product, it also provides multilayered representations for signals, using reconstructions from a given window type only