On Tue, Jun 22, 2010 at 8:00 AM, Adrian Knoth
<adi@drcomp.erfurt.thur.de> wrote:
On Tue, Jun 22, 2010 at 06:38:45AM -0400, Gene Heskett wrote:
> >from TFA:
> >: Implemented in a DSP chip or microprocessor, this simple compressor
> >: requires about 50 instructions per sample. However, lossless
> >: compression ratios fall between 1.3:1 and 2:1 on baseband signals.
> >
> >So a size of 75% expected and on occasion down to 50% after compression.
> >How is that compared to existing implementations?
> It was the lossless claim that got my attention, Jens. I am well aware
> that current compressors can beat that at "acceptable" quality. But an
> ogg at q7
Jens was never talking about lossy "compression", which I call data
reduction to avoid the ambiguity with real compression (as in ZIP).
Lossless audio coding is nothing new, FLAC has been around for years.
Your referenced codec achives 1.3:1 to 2:1. One can compare this to some
values provided here:
These are all lossless codecs, and as one can see, only few manage to
come close to 2:1 (50% compression).
However, results for predictive coding (derivation based approaches)
vary a lot depending on the input signal. As a rule of thumb, a pure
sine is easier to predict than noise, which more or less is the
mathematical equivalent of randomness (I'm sure Fons could go into
detail here, if necessary).
Long story short: I don't think your link contains something
extra-ordinary, just another me-too approach of well-known techniques.
It might save you a few bits, but you'll have to measure it. Fire up
octave, load the matlab script, encode a wave file and compare it to
FLAC.
If your referenced algorithm gives striking results, then convince
everybody to forget about FLAC and use this new algo instead. Let me
predict that neither the first nor the latter will happen. ;)
That's more or less the end of the story. Any further discussion would
only make sense with measured results at hand.
HTH
--
mail: adi@thur.de http://adi.thur.de PGP/GPG: key via keyserver