The digitally sampled data is split into a plurality of subbands each covering a different frequency range. The subbands are each individually expanded to normalize the energy in each band and the subbands are converted by FFT to the frequency domain and the magnitude and phase portions are processed by different techniques based on psychoacoustic principles. Magnitude data are processed by tree structured vector processing to develop code books for each subband which are unique to each song. Phase data are uniformly quantized with dynamic bit allocation used to increase resolution on transient passages.