[vorbis] Vorbis determined to be as good as MPC at 128 kbps!

[Segher Boessenkool]

Hi Sebastian, list,

>>> The HF boost is a modern example.  Sebastian's paper correctly
>>> identifies one major source of the slight HF boost people noticed in
>>> Vorbis some time ago.
>>
>> Using some perceptual models, yes.  Using other models, not at all.
>
> ?!

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

The only "real" test is a listening test, of course.  As far as
I know, no-one did one with your proposed modification yet.

>>> The solution he suggests is also perfectly workable.
>>
>> Workable, but that doesn't mean anything.  A simple noise gate
>> would work even better!  (Using his model.  I haven't spent the
>> time necessary to evaluate if it actually sounds any better).
>
> I think it's worth mentioning that I agree with Monty, a
> modification of the noise normalization code to compensate for the
> boost effect is a better idea compared to the solution I gave.

And so do I.  Or, a little stronger: a modification of the noise
normalization code _to not exhibit this boost effect_.

> (Basically the effect is pretty much the same but the NN sheme
> does not rely on a certain distribution and adjusts the residue
> samples dynamically)

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

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

> I'm not sure about what you mean by "noise gate" and how this is
> supposed to be superior to an improved NN sheme.

I didn't say that.  I said that it would be superior to your
proposed scheme.

A noise gate filter is one of the traditional old-fashioned
audio filters.  In its simplest version, it just clamps all
signals close to zero to zero.

Your scheme offsets all signals by a fixed distance into the
direction of zero, before quantization; a noise gate does the
same, but only if the resulting value (after quantization)
becomes zero.


Hope this makes things clearer,


Segher

p.s.  Followups to vorbis-dev?