> But if that conversion were to floating point rather than fixed
> point data, and then converted to fixed point after the attenuator,
> that would seem to be a valid way to get the signal in the right
> gain range for storage in a file.
I think a simple digital attenuator would not require floating point arithmetic
at all because there is only a single multiplication involved. Things may
become a bit more complicated if an additional 24 to 16 bit conversion with
dithering was required, but I guess that this could still be done properly with
fixed point arithmetic.
I also believe that most (if not all) digital field recorders currently use
low-power fixed-point DSPs rather than more complex and more power-hungry
floating-point models.
Post processing in sound editing software however would be a different story
because such a system must be prepared to properly handle numerous subsequent
processing steps where rounding errors could be an issue.
Interestingly, even the 32 bit floating point format can potentially cause
trouble in algorithms that require accumulating many small numbers into a much
larger sum.
Regards,
Raimund
|