Menno wrote:
> While implementing M/S threshold calculations for FAAC, I noticed some
> things that could be improved in GPSYCHO.
>
> The formula for the BMLD protection ratios you reverse engineered should
> look like this:
> mld[b] = pow(10, -3*(0.5+0.5*(M_PI*(min(bval[b], 15.5)/15.5))));
> I took this from the 14496-3 ISO document.
> If you want to see what the rest of the calculations should look like, you
> should take a look at psy_step11andahalf() in psych.c from FAAC 0.55beta.
> Also the paper advises to use the more tonal tonality values from the left
> and right channels for the mid and side channels. And the minimum of
> unpredictability from the left and right channels for both the mid and side
> channels.
> I haven't tried this in LAME myself, but in FAAC it works very good.
>
> Bye, Menno
Hi Menno,
I just compared your mld with the one implemented in Lame, and they
are totaly different, even in the tendency.
With your mld = -3*(0.5+0.5*(M_PI*(min(bval[b], 15.5)/15.5)))
and Lames mld = 1.25*(1-cos(PI*b/SBPSY_s))-2.5
the one in Lame increases monoton, yours falls monoton.
Here are the mlds for short blocks:
Lame mld your mld difference
-2.5 -1.5 -1
-2.45741 -2.02384 -0.433572
-2.33253 -2.54737 0.214835
-2.13388 -3.03745 0.903571
-1.875 -3.45245 1.57745
-1.57352 -3.81333 2.2398
-1.25 -4.14806 2.89806
-0.926476 -4.43749 3.51101
-0.625 -4.69682 4.07182
-0.366117 -4.93336 4.56724
-0.167468 -5.15104 4.98357
-0.0425927 -5.33832 5.29572
So, what is going on here?
And where can I find the sources for FAAC 0.55beta? I only found
the sources for FAAC 0.42.
Robert
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
