An elegant coding scheme is described that results in more than 120:1 compression ratio of the speech signal. Linear contractive and affine transformations are applied to the difference between the FFTs of two consecutive speech frames. A set of rules that represent this difference are deduced and transmitted or stored. These rules are used to reconstruct this difference and ultimately the speech signal itself. The reconstructed speech is superior in quality to LPC speech. Also this scheme is well suited for the hypermedia coding and data compression problem.