We propose a signal based joint time-frequency segmentation algorithm as an extension to Herley et al. (IEEE Trans. Image Processing, 1997). Our algorithm provides an unconstrained multiresolution analysis in time and frequency adapted to the characteristics of the signal, where the time-frequency segmentation of Herley et al. is modified to achieve a minimum entropy criterion. Experimental results on a synthetic signal, composed of a high frequency sinusoid and a single impulse, show that our algorithm outperforms several time-frequency representations such as the best basis wavelet packet, the best basis modified discrete cosine transform (MDCT), and the original Herley et al. algorithms. The application of our algorithm to the transient information in the signal provides good results and shows that the algorithm will be useful in speech decomposition problems. Copyright © 2008 ISCA.