To follow up on Greg's VBR comments:
I agree. The perceptial model is just a crude guidline, and using
it with VBR (which will find scalefactors to get the noise just
up to the limit of the allowed masking) reveals a *lot* of flaws
in the psy-model.
I just put lame3.33 on the web page which I think is a big improvment.
I've just done a few simple tests, and the a VBR average
bitrate=128kbs not obviously worse than the fixed 128kbs. It uses
Robert's #define RH code (which is now the only VBR code) and always
quantizes to 0 audible distortion. The quality is improved by
lowering masking from the psy-model output, by up to 10db at VBR_q=0.
And the masking is further reduced (by up to 6db) as the bitrate
decreases (so you need an extra 6db of 'quality' to get all the way
down to 32kbs).
The one thing VBR should be good at is a replacement for the bit
reservoir. For something like castanets.wav, the bitreservoir is
completely drained after a few castanet clacks. With VBR, you can
increase the bitrate whenever needed. Thus I like the idea of
having a 'target bitrate' as Greg suggests. We just replace the
bit reservoir with the more flexible VBR - increasing bits when
needed and allowing for small decreases when the average bitrate
gets too large.
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
