>
> Looking at the ISO docs annex 3-D.2 Psychoacoustic Model II:
>
> 4. Calculate the unpredictability measure cw
> ...
> By sacrificing performance, this measure can be calculated
> on only a lower portion of the frequency lines.
> Calculations should be done from DC to at least 3kHz and
> preferably to 7kHz. An upper limit of less than 5.5kHz may
> considerably reduce performance from that obtained during
> the subjective testing of the audio algorithm.
> ...
> Best results will be obtained by calculating cw up to 20kHz.
>
> does this mean, we can improve GPSYCHO by calculating cw up to 20 kHz?
> Now we do it up to 8.9 kHz I think, and how much performance will we loose
> doing so?
>
Hi Robert,
Please read page 80: 'adaptation of psychoacoustic model II for layer
III', it has a section on calculation of the unpredictabilty, and
those recommendations are what LAME uses. The changes you are
suggesting come from the layer II documentation. I would bet that the
predictability measure is rather meaninless for the high frequencies
(it measures the correlation between FFT coefficients from each of the
three short blocks) and so using a value of .4 is better than the
estimate coming from the tonality formula.
If you want to change this, please do not just check your changes into
CVS, since this breaks my validation test cases, and I will just
undo the changes. To make a change like this (which has uncertain
benifits) you should make your own CVS
branch and then find a test case where the change does improve things.
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
