[vorbis] Re: [vorbis-dev] high frequencies response

[Monty]

On Wed, Oct 09, 2002 at 08:32:02PM +0200, Frank Grotelueschen wrote:
> Hi there
> 
> In the past, i have used lame to encode high quality mp3-files
> (vbr 1, bitrate ~ 192kbit). I tend to switch to ogg with Quality 4
> or 5, but i noticed, that many ogg-files tend to produce too much
> high frequencies response. In many cases, this is very noticeable.
> For my opinion, i cannot accept this worse frequence reponse.

I've been assuming for some time that this phenomenon (high
frequencies seeming amplified) has been a combination of two things:

1) When point stereo rotates noise audio into the middle of the
coupled image, the energy does not increase (the rotation is correct
and elliptical), but because the sum total of diffuse noise is now
focused in one place, it is more immediately noticable.

2) Quantization of very low level energy, unlike mp3, does not
purposely minimize added quantization noise.  It either allows the
energy to creep up, or if things quantize to *much* lower than
original energy, will purpousely artificially support it.  The phase
and presence of noise energy is much more important to the charachter
of the noise than minimizing the naieve vector difference summation
ala mp3.

Both of these contribute to the impression of boosted highs.  Rather
1) gives the impression, but 2) can actually result in a few dB of
boost.  This is why it seems like at -q 6 the problem mostly
disappears, when lossy stereo is finally completely switched off.

Note that these are actually very reasonable sounding theories I've
only partially tested and justified to myself.  They're not bugs so
much as shortcomings in the encoding psy model we'll need to correct.

Two steps here: First verify these are in fact the causes, then fix
them.  1) is hard to deal with as it's somewhat inherent in low-level
point stereo.  2) is relatively easy to fix conceptually, what remains
to be seen is how annoying it will be to code.

Most of the tests people have supplied to 'prove' what is going on
have been making incorrect assumptions about the encoding model and
thus are flawed in one way or another.  I'd not mind gathering a few
of the people who are *really* sensitive to this boost (I'll admit
that I notice it, but it doesn't bug me) to run tests verifying what
is going on.  Frank, are you up for this?

Monty
--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to 'vorbis-request at xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.