>
> Mark, would you give me more detailed explanation about this?
> How can I reproduce this?
>
> Mark> For example, two identical signals, just
> Mark> phase shifted and with different low-level noise, can generate very
> Mark> different maximum values,
>
> --
> Naoki Shibata e-mail: [EMAIL PROTECTED]
>
Before today, I had never looked at this in MP3 - I was
just spouting off based on general Fourier series experience from
my day job :-)
In MP3, a good example is "60.wav", which IIRC, you created :-)
It is a higher frequency tone superimposed on a 60Hz tone.
The only difference in the signal from from granule to granule is the
phase of the two frequencies. But take a look at the energy in MDCT
coefficient 2 (left channel), for example, in frames 10-20, varies
from -16 to -47 db. So the uncertainty in this single MDCT
coefficient coming just from phase errors in 31db!
The usual DSP techniques to reduce this uncertainty:
1. take a larger MDCT window (not possible with MP3)
2. Smooth the spectrum. If you instead take the average
of MDCT coefficient 1,2,3 you get values from
-13 to -23db. So the uncertainty is down to 10db.
I repeated this experiment for MDCT coefficient 26 (near the
second spectral peak in 60.wav). The values for frames 10-20:
MDCT coefficint 26: -38db to -19db Variation: 19db!
average of 25,26 & 27: -26db to -24db Variation: 2db
I didn't look at the energy computed by the FFT, but
the FFT is supposed to give better energy estimates.
(which is why it is used in the psycho acoutics instead
of the energy values computed by the MDCT)
How does this relate to calc_noise/MAXNOISE issues? I'm not really
sure, but I just wanted to make the point that in the code,
you *cannot* trust data coming from a single MDCT coefficient
(or a single FFT coefficient).
Mark
varies from
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
