> There are still several ways to improve CELT, such as by incorporating
better
> psychoacoustic masking in the dynamic bit allocation. This is a difficult
> problem both because there are few bits available for coding the
allocation
> and because the analysis window is short.>
> it is possible to implement that and what are the correct approaches to do
that.
Several people have tried, including Garf and Jean-Marc. No one has been
able to demonstrate an improvement, even when you don't count the
overhead required to signal the allocation, i.e., even when you flat-out
cheat. This does not mean it can't be done, but it does mean we don't
know how to do it.
We have considered adding a method of coding a boost to one or two bands
(e.g., reusing the PVQ code), which is signaled with low probability.
That would have little impact on the normal case where it is not used,
but the lower the probability, the bigger the gain you'd have to get
from your psychoacoustics before it would be useful. The other drawback
is that if the feature is not used in practice, it will not be
well-tested, and thus unlikely to work if you ever do figure out how to
use it.
If anyone _can_ identify a method of doing dynamic allocation that makes
a consistent improvement, that would be very useful. We are hoping to
finalize the bitstream in the near future, though, so there isn't a lot
of time left to try.