[FFmpeg-devel] donation for snow
Fri Nov 7 09:47:29 CET 2008
> Please be more specific, there are people here trying to write a decent
> AAC encoder.
I've already mentioned a lot of this in #ffmpeg-devel, but x264's
ideas, translated into audio, are twofold:
1. Sections with lower complexity need higher relative quality than
sections with higher complexity. That is, a very complicated segment
of sound will naturally mask artifacts much better than a simple
segment, and additionally simple segments tend to have smaller
frequency values, meaning that the quantization error is higher, e.g.
when quantizing a value of "1.5" you will lose far more relative
accuracy than quantizing "9.5". This is partially offset by AAC's
x^(3/4) quantization method.
2. The energy of each band should remain similar to the original
energy. This method is used in both Vorbis and Xiph's upcoming
low-latency format Celt, I believe. Celt, in fact, explicitly codes
the energy of each band into the bitstream itself.
More information about the ffmpeg-devel