> From: Takehiro Tominaga <[EMAIL PROTECTED]>
> Content-Type: Text/Plain; charset=us-ascii
> Date: Wed, 24 Nov 1999 00:22:30 +0900
> X-Dispatcher: imput version 990905(IM130)
> Sender: [EMAIL PROTECTED]
>
> Hi all!
>
> I don't know Johnston & Ferreira 1992 ICASSP paper,
> but I think this is so strange code....
>
> in psymodel.c, we adjust the MS threshold if L & R masking differs
> by 2db or less.
>
> the code is
>
> if ((chn==1) || (ratio[0][sb] <= 1.58*ratio[1][sb]) ) {
> :
>
> This condition is not symmetric. I think "differs by 2db or less" should not
> mean "diff < 2db", but "-2db < diff < 2db" from the symmetry of LR channel.
>
Yes indeed, another bug (my carelessness).
> So I made this patch. please check this out.
>
> >>>>> "G" == Greg Maxwell <[EMAIL PROTECTED]> writes:
>
> G> Has anyone looked into the 'left channel attentuation' claimed
> G> by the USE mp3 test site?
>
> I have not tested this patch very much, but I hope it will help the problem.
This bug will effect the mid and side channel masking. What is
supposed to happen is that if the L & R maskings are about the same,
then the side or mid channel masking is allowed to mask both
the side and mid channels. (except for some high frequency
correction which I never understood) The bug will allow
this interchannel masking even if the L is much weaker than
R.
A more serious problem is a few lines down: The MS switching
criterion has the same mistake. I compute the average
difference between the L and R masking, but again I didn't
take into account that the log of the ratio (in db) can be negative.
Changing this will require retuning the MS switching theshold,
which is a rather tedious process.
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
