--- In "Steve Pelikan"
<> wrote:
> Dan Dugan, in a reply to Klas Strandberg (about digital
manipulation, and
> especially w.r.t normalization), wrote
>
> > In intermediate steps, you need headroom
> > for processing.
>
> My question: I've always assumed that most digital manipluations
were done
> as floating point (thus giving
> more headroom than even long integers could require.) Is this
true? If not,
> what are the problems with using
> floats?
>
> In the sound manipulation program (1) I'm working on, I've
temporarily
> provided means to perform manipulations either with either integer
or
> floating point aoperations because I didn't know what was
considered
> superior. I can't say that I've ever noticed a difference.
>
> By the way, I don't normalize either.
You are right, one would be always on the safe side when using
floating point formats for digital manipulations. For instance, if
you rose the level temporarily beyond the initial dynamic range of a
sound file, you would introduce clipping to an integer format file.
Instead, a floating point file would maintain all the information
without clipping.
As long as you do not rise levels, there would be no significant
differences between integer and floating point processing. But
excessing repetitive volume manipulations in integer files
(especially when you first decrease and then increase levels) may
lead to additional noise or distortions. Though, common (FIR)
filters and equalizer using integer arithmetic should work
satisfying too (in critical internal stages, they use longer integer
numbers of 32 or even 64 bit).
I would also try to avoid any excessive normalizing. Imagine, you
had a very soft recording of a bird whistle with a very low noise
floor (perhaps as a result of applying a noise gate filter). If you
then raised the level of that whistle (a relatively pure sine wave)
by e.g. 20 dB you would get a very unnatural result. That artificial
amplification would add spurious harmonics to the whistle, because
there is no information on the original fine-structure of the
waveform. The software has no chance to perfectly reconstruct the
original pure-tone signal at higher levels (it does not know,
whether the original sound had softer harmonics or not). Usually,
these harmonic distortion will be inaudible, because it would be
buried deep in the noise floor of the recording. However, if the
original noise floor has been removed by applying a noise gate
filter (or even by the inherent noise gate effects of ATRAC / MP3
compression algorithms?), this kind of distortion could become
audible.
Regards,
Raimund
Avisoft Bioacoustics
http://www.avisoft.info
________________________________________________________________________
________________________________________________________________________
|