Combined prediction and residual coding for lossless audio compression

Summary form only given. We propose a new paradigm for lossless audio compression, where prediction and residual coding are combined in a single logical stage, taking advantage of the fine statistical structure of the original signal. We obtained significant compression gains for integer signals which were normalized, companded, or recorded using DAT-LP. We compared our existing lossless audio compressor OptimFROG and the new modified variant on a large corpus of 73 Audio CDs. The encoding time was increased with 9.6% and the decoding time with 6.1%. Overall, OptimFROG (normal) obtained a compression (compressed size/original size) of 55.96%, the new variant 55.36%, and Monkey´s Audio (high) 56.10%. For 19 CDs we obtained compression improvements up to 12.35% and on average 2.14%, for 14 CDs we obtained small compression improvements on average 0.24%, and the remaining CDs did not present improvements. For mu-law decoded, A-law decoded, and DAT-LP audio, we obtained on one file compression improvements of 28.66%, 22.56%, and 20.98%, respectively. The new paradigm can be also applied to 24 bit lossless audio compression and to lossless image compression (further favored by the small bit depth per color)