[FFmpeg-devel] [RFC] AAC encoder optimizations

Gabriel Bouvigne bouvigne
Sun Sep 14 16:17:54 CEST 2008

Kostya a ?crit :

> Since I can work on it mostly on spare time during work I would be grateful
> if you suggest some tests for quality. 

When working on bit allocation or quantization stuff, you should NOT 
rely on any listening test. As you don't know yet if your psymodel is 
good or not (and anyhow it will never be perfect), it should not play 
any part within development of bit allocation.

Instead, you have two possibilities:

*in bitrate constrained cases, you should rely on SMR values(signal to 
mask ratio) computed by your encoder. Improvements in bit allocation 
must provide in increase in SMR values.

*in non bitrate constrained cases, you should realy on final size for 
the same SMR value, ie for the same target SMR improvements in bit 
allocation must reduce number of bits. (targetting transparency, ie 0 
distortion, might make you job easier for such tests).

By removing dependance on psymodel accuracy (ie getting rid of listening 
tests) when working on bit allocation stuff, you will make your job a 
lot easier. If you try to test bit allocation (potentially buggy as it 
is still in development )by using listening tests, you will never know 
if the issues are within bit allocation or psymodel, and you will 
quickly loose your mental sanity.

Moreover, when relying purely on mathematical results when improving bit 
allocation, you are then able to automate tests and to run tests on a 
lot of data.


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

More information about the ffmpeg-devel mailing list