Re: [Lame-dev] block size detection

[Gabriel B. <bou...@mp...>]

Robert Hegemann a =E9crit :

>>*There is an highpass filter, then energies of subshorts are computed.
>>But the contribution to energy is not frequency dependant. It seems to
>>me that an high energy located at 18kHz would have the same contributio=
n
>>as an high energy located at 12kHz, while the former is not necessarily
>>audible.
>=20
>=20
> so you are proposing to weight energy by some audibility factor?


I do not know if that should be a weight or a noise gate according to=20
ATH, but I think that energy surge on inaudible components does not lead=20
to smearing of quantization noise. The problem is that such=20
frequency-based filtering has to be done in the frequency domain, while=20
it seems that this high pass is done in the time domain.

This brings another observation:
 From my understanding, the high pass filter is done on input time=20
domain sample, which are not yet lowpassed. That would mean that high=20
freqs that will be lowpass filtered (according to the encoder settings)=20
are used for the block size selection.
Perhaps there should be a lowpass there.

...or this detection could be done in the frequency domain, but in this=20
case using sub_short blocks requires additionnal computations.


--=20
Gabriel Bouvigne
www.mp3-tech.org
personal page: http://gabriel.mp3-tech.org