Hi,
Before the former -X n experimental settings in quant_compare are dropped
entirely, I'd like to report about the experiences I made with these
evaluation heuristics.
1) experimentalX=0 uses the most stable setting (and the default method).
2) experimentalX=6 (Robert's?) achieves the highest quality (for my
liking). The sound seems to be clearer and more crisp. This mode also cooperates
well with the high quality settings -q 3 to -q 0. (-h is identical to -q 2).
3) stabilty of experimentalX=6 is not optimal. I think this is the result
of normalizing over_avg_noise to the number of bands with noise above
threshold (over variable in calc_noise1). I don't think this is good.
Normalisation to the number of bands instead is more stable and shows no regress in
quality:
- res->over_avg_noise = over ? (over_noise / over) : over_noise;
+ res->over_avg_noise = over ? (over_noise / count) : over_noise;
over even: res->over_avg_noise = over_noise/count; (as far as count!=0).
The new CVS versions use the -X switch for setting VBR_max_bitrate.
To activate the old -X n settings for testing, initialize
gfp->experimentalX = 6; (or other desired value) in lame.c.
Regards,
Alfred
--
Sent through GMX FreeMail - http://www.gmx.net
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
