> X-UIDL: I:m"!@9G"!KPD"!i2'#!
>
> Mark Taylor wrote:
> >
> >
> > To explain how this works, take a 128kbs CBR for example.
> > In that case, LAME allocated a base amount of
> > bits for each frame. This base amount is about 10% less than
> > a 128kbs stream, so that the bit reservoir is slowly built up.
> > LAME then allocates extra bits based on the "perceptual entropy",
> > computed in psymodel.c. If the perceptual entropy is large,
> > the frame can use the base amount + up to 60% of the reservoir.
> > These ideas come straight from the ISO docs, but the implementation
> > in LAME is new (the ISO dist10 code reservoir treatment is a disaster.)
> >
> > "Safe VBR" is even simpler: It computes the same base
> > amount of bits for each frame, and then computes extra bits
> > needed based on the perceptual entropy. But instead of
> > relying on the bit reservoir (which might be empty)
> > it just increase the frame size when extra bits are needed.
> >
>
>
> Hello,
>
> Isn't this "Safe VBR" (let's say with 128kbs as a base bitrate)
> equivalent to a classical VBR with a minimum bitrate: lame -v -b 128
>
> ??
The is one key difference: safe VBR mode chooses the total number of
bits to use before *any* quantization occurs, based on a simple
formula (also used by the CBR algorithm). The psycho acoustics are
only used in allocating the bits among the frequency bands.
The other VBR modes decide on the total number of bits to use based
soley on how the quantization errors compare to the psycho acoustic
masking. Ideally, this is a better way to do things, if your psycho
acoustic model was perfect.
But the psycho acoustic model in LAME far from perfect, and trusting
it as much as VBR does can lead to mistakes: A good example of this is
"vbrtest.wav", which was just added to
www.sulaco.org/mp3/gpsycho/quality.html
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
