> X-Authentication-Warning: cs.csoft.net: $s=geek.rcc.se doesn't match
>$[EMAIL PROTECTED]
> X-Authentication-Warning: geek.rcc.se: majordom set sender to
>[EMAIL PROTECTED] using -f
> From: Robert Hegemann <[EMAIL PROTECTED]>
> Date: Sun, 21 Nov 1999 11:19:20 +0100
> Content-Type: text/plain
> Sender: [EMAIL PROTECTED]
> Precedence: bulk
> Reply-To: [EMAIL PROTECTED]
> X-MIME-Autoconverted: from quoted-printable to 8bit by cs.csoft.net id FAA27622
> X-UIDL: DGHe9nh\d9TBf!!cp:!!
>
> Mark Taylor wrote:
> > This sounds like a good idea. But if the thinking is correct, then
> > I dont see why we should ever use the average noise. How about
> > if we just replace ave noise with max noise?
>
> Well, if you replace average noise with maximum noise, then you will
> end up in LARGE vbr files. And if there is not so much difference
> between average noise and maximum noise, then our psy model works
> just fine. Only these strange situations when the difference is large,
> because we have one outstanding frequency line within a band,
> make Lame problems.
>
My main problem with the patch is that is basically a complicated way
to do a weighted average (between ave_noise and max_noise), with TWO
parameters that must be tuned. And what is the meaning of 1e-7?
I'm still leaning towards just replacing ave-noise with max_noise.
When max_noise >> ave_noise, this will be about the same as your
patch, and when there is a small difference is does not matter which
one you uses. But I did a few tests at fixed 128kbs: the differences
were very minor, but mixed (some improvements, some degradations).
> Maybe this is it:
> * why side channel sparsing didn't work?
> * why we make some extras at low bits? (increasing db_lower ie.)
>
I still think sice channel sparsing is a red hearing! it is only
mentioned in the ISO docs, none of the technical papers ever mention
it.
I also dont like increasing db_lower for low bits!
When it is not used, it makes it easier to
test how accurate the psymodel is by using -V5, since then -V5
means quantize so that there is all noise is < allowed masking.
Try the new spreading function (#define NEWS3 in psymodel.c) with -V5
also. It is the simple version from the Painter article, and overall
represents *less* masking then the default spreading funciton,
resulting in higher quality (and bigger files) when used with -V5.
This weekend I did a lot of tests comparing with and without NEWS3 at
128kbs fixed bitrates, but again I couldn't really decide what was
better.
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
