The idea came to me during a meeting about multifractal analysis of clinical EEG data with my professor, but he mentioned this guys name... and paper.
R.A. Wannamaker and E.R. Vrscay, Fractal Wavelet Compression of Audio Signals, Journal of the Audio Engineering Society, Vol. 45, Nos. 7-8, pp. 540-553 (1997).
Now i know people are working on different kinds of wavelet compressors, but haven't seen anything published that shows them having better compression (or coming close) as the current lossless formats [for cds].
Now, i am not a professonal in audio codecs design but from my limited understanding, looks promissing..... just wanted to see some "professional" opinion (or anyones for that matter)
Is it worth pursuing?
It all depends on the shape of your signal, and how fast you can approach it with what you have.
For a pure tonal signal a Fourier or (M)DCT is very good, and probably for more agressive sounds the wavelet approach is better.
The impressive wavelet results I've seen, are in lossless compression of photographic images. For sound, it has to be tested more thoroughly.
If you want to have some fun, go ahead ! If I had a bit more time and more experience in wavelets I'd do it too !
Compression ratios above 6:1 should ultimately be attainable with good fidelity signal reconstruction
The theory seems promising INDEED, usable in lossless compresors as well as in lossy ones. Pity my lack of math knowledge, I can't understand as much as I wish.
EDIT - Upppssss. The paper is from 1997. IMO anything worth to test has already been tested and discarded. Not so promising
The impressive wavelet results I've seen, are in lossless compression of photographic images.
Really? Could you share a link?
I saw impressive wavelet results only in lowbitrate area. Best lossless results I saw were still worse than of JPEG-LS.
lossless jpeg2000... works better than lossless JPEG. typically it'll give a slightly larger result to lossy JPEG at "full quality" (like q 100 in photoshop).
search for the "kakadu" jpeg2000 encoder. aptly named after a vast wilderness with too much uranium for it's own good
At the last SCALE I talked a little with Monty about this. He has some ideas for using wavelets in lossless coding but not much time to try them. Ditto for me.
JPEG2000 lossless beats lossless-JPEG, but on average it under-performs JPEG-LS. The state of the art lossless image coders generally dont use transforms.
See http://itohws03.ee.noda.sut.ac.jp/~matsuda/mrp/ (http://itohws03.ee.noda.sut.ac.jp/~matsuda/mrp/)
All in my opinion :
Quantization reduces intercoefficient correlation, that is why transforms work well for lossy coding. With lossless coding you cannot afford to ignore the intercoefficient correlation, but when you stop doing that the transform really isnt worth it anymore ... it actually makes modelling harder.
On the other hand music has more long range repetition than images ... but that cant really be efficiently used by fixed transforms AFAICS.
Some type of matching/basis pursuit with online training (slowwwwww) followed by lossless coding of the residual might work. You could try cutting up the music a couple of times in sections who's size is determined by major peaks in the autocorrelation function, then for each size applying PCA or something to get "atoms" for basis pursuit. That's getting really really complex though.