[music-dsp] 32 bit floats >>> 24 bit integer

[robert bristow-johnson]

> > Despite these quirks (truncation in floating point versus integer and 
> > various types of rounding possibilities) that make it more difficult to 
> > directly compare truncation versus rounding, your assertion that rounding 
> > isn't a significant improvement over truncation is correct-- they are 
> > simply different methods of spreading the conversion error.
> 
> I disagree.  Clearly the type of truncation that goes from something like
> 5.9 to 5.0 is going to produce a statistically larger magnitude of error
> over time then rounding from 5.9 to 6.0.  Anytime the truncation exceeds
> 0.5, it is worse.  Plus it's always biased in one direction.

but the component of that error that makes it bigger is simply a DC offset.  particularly if this quantized signal is getting fed back somehow, this DC offset can grow and become an annoyance and possibly cause asymetrical clipping, but otherwize that extra error doesn't sound like anything.  the statistics of the AC component of the error for "round toward -inf" are the same as "round to nearest". 

--

r b-j                  rbj at audioimagination.com

"Imagination is more important than knowledge."