Thank God, audio signals tend usually to be quite different from pure white noise. The method described in this paper takes advantage of that basic fact in converting linear PCM to a new representation of audio content. No magic is involved but the simple application of some general information theory basics. This paper describes some fundamentals on how to decorrelate music or speech samples by the use of predictive coding technique involving a multi-stage approach, optimized for the requirements of a real time implementation. Some implementation examples are then examined both on general purpose CPUs (such as IBM PC) and on DSP. Resulting charts are given to illustrate the compression ratios obtained for various types of audio signals in comparison with the entropy of the computerized signals.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.