The Logic Off Topic list

Thoughts from the mind of monsdrum, 22-02-2002:

>  > Normalising obviously introduces some distortion -- the numbers
>  > have to be multiplied by some factor and then rounded to the next whole
>  > number --
>
>Afaik this isn't right. To my understanding, the numbers are not
>multiplied by a factor, but rather added to a number high enough to
>make the peak hit 0 db. Thus there is no rounding anything, and no
>distortion.

Huh?  So you say that if you have e.g. (silly example) 2 samples with 
values 1 and 2, and normalise them to the full 65536-scale (16 bit), 
you would get samples with values 65535 and 65536?  This can't be 
true, as the relative volume of both samples isn't preserved.  Those 
samples should become 32768 and 65536 -- i.e. multiplied by 32768, 
and thus still a factor 2 apart after normalising.
Hence my idea that some distortion is introduced: if you have to 
multiply by some non-whole number, such as 1.3, most resulting values 
will be non-whole and thus have to be rounded.

[5 minutes later]  OK, I just did a simple experiment: take an audio 
file and draw 2 humps in it, one going up to 10% and the other to 
20%.  Select and normalize.  The result is 2 humps at 50% and 100%. 
So yes, there's definitely some sort of multiplication involved.  The 
same result can't be achieved by whichever kind of addition...

>The problem is rather, that when raising the peak level you also
>raise the noise floor.

Sure, that would be another problem.

-- 
Hendrik Jan Veenstra  <h@...>
Omega Art: http://www.ision.nl/users/h/index.html