[Vorbis-dev] Re: [vorbis] Vorbis determined to be as good as MPC at

[Sebastian Gesemann]

128 kbps!
In-Reply-To: <E77B14D5-B72A-11D8-91A8-000A95A4DC02 at kernel.crashing.org>
Message-ID: <Pine.LNX.4.44.0406081019270.6447-100000 at gorlois.cs.upb.de>

On Sat, 5 Jun 2004, Segher Boessenkool wrote:

> There are quite a few models of how the ear works.  All the good
> ones are computationally expensive, and not usable at all
> mathematically.  Your paper uses just one simple model.  Using

model of what ?
ok, it models the coefficient distrubution with a gaussian
curve. check for yourself. the distrubution IS gaussian-like.
the solution just tries to keep the energy of the quantized
signal close to the energy of the original signal which
'plain uniform quantization' does not (as showed).

the desire of keeping the energy level of the quantized signal
as close as possible to the original one is of course reasonable
because the perceived loudness depends on objectivly measurable
energy.

i'm not saying the solution i gave is perfect. (it would be
close to :) if the coefficient distribution perfectly
matches the gaussian model)

> some other models, the HF boost is not explained by what is in
> your paper.

can not comment on that because i've no idea of what you
are talking about. you probably misused the term 'model'.

"paper" ? - it's badly written and i'm aware of that :)
i just wanted to publish my results quickly without
spending too much time on correcting spelling errors
and stuff. if i had to rewrite it, it would be filled
with more details, easier to understand and twice as
long probably.

> The only "real" test is a listening test, of course.

the whole "the real test is a listening test, of course"-thing
is a well-promoted justification.

it's for people who think they can determine the perceived
quality of a compressed song by looking at Cool Edit's
spectrum view and stuff...

a listening test is a test for measuring _what_ ?
to measure if the SNR / low-pass is too low so that differences
can be perceived ? - YES!
to measure if the loudness (energy) has changed ?
- CAN DO as they are some sort of 'difference'

but the quantization energy boost can be avoided without the
need for listening tests here.

> I'm not a fan of the noise normalization thing, I have to admit;
> because it's a) a heuristic (not based on a model) and b) I don't
> understand why it works.  a) /an sich/ is not a problem, but
> combined with b), I don't like it ;-P

as i said, an improved NN sheme will just work like the solution
i gave. the only difference there will be is that the threshold shift
'd' will be determined adaptivly based on the signal and an analysis-
by-synthesis approach.
So, it does NOT model the coefficient distribution with a gaussian
curve and use fixed shifts. it's therefore similar but more
powerful in case of non-gaussian distributed coefficients.

It's IMHO NOT heuristic.

> It would be great if someone could explain it to me.

the original NN sheme checks the energy of residue partitions
of size 32 (long) and 8 (short). in case the energy drops as
an effect of the quantization the samples which caused the
heaviest energy losses are altered thus increasing energy
and approximating the original level.

the NN sheme in 1.0.1 just increases energy in case it
dropped. usually quantization results in an energy boost (for
SNRs above 2 dB). but this situation is not handled
by the current NN sheme because this effect was unexpected.


Ghis!
Sebastian

--
PGP-Key-ID (long): 572B1778A4CA0707