The research efforts recently devoted to the theoretical treatment of time-varying filter bank structures have created an adequate environment for the development of adaptive analysis tools, useful in many signal processing problems. In this paper we present a complete adaptive analysis system suitable for the compression of audio signals; we would like to remark that our goal is not to develop a complete audio coder, just the appropriate analysis/synthesis scheme. For basis selection, we propose a search algorithm and a new cost function based on perceptual considerations. In addition, a procedure that determines the optimum length of the analysis segment is described. Examples are given of the performance of these algorithms on some representative audio signals. When the search method is compared with previous basis selection algorithms, we find and show that, for the application of audio compression, the perceptually driven basis selection algorithm leads to a much higher compression efficiency and better quality of the reconstructed signal.