The main reason you should not normalise is quantisation and rounding errors.
If you recorded say 16 tracks of audio and normalised them all, then went about mixing them together in a digital mixing app using faders to get a mix, What you have done is level quantised the bits in each audio file twice!
Moving a fader in a digital mixer is just like normalising but your aim is not to turn it all the way up to zero.
More on quantisation.. (this explanation will be using a hypothetical 16bit digital mixer)
say you have recorded a cute sine wave at -12db off hitting zero. You did this in a 16bit file, for each -6db away from zero you record your using one less bit.
a 1bit file would contain 2 states, silent and loud as hell.
a 2bit file would contain 4 levels (2x2), silent, soft, medium loud and loud as hell.
a 4bit file would contain 16 levels (2x2x2x2),, and it goes on like that
...
...
...
a 14bit file would contain 16384 levels
and a 16 bit file would contain 65536 levels
now thats a 1 to 4 ratio, Your volume steps in your 14bit file are mapped out to fit into the space of 4 steps for every one step. Thats great if your little sine wave is exaclty 16384 levels loud, if its say 16001 levels loud (still a 14bit number and still -12db) there will be not exactly a 1 to 4 ratio, rather 1 to 4.0957440159990000624960939941254 ratio.
That's when rounding errors occur because the final level has to be a whole number without the .0957440159990000624960939941254, it will round it up or down, or what ever the software designer has in mind.
Then you go and turn your track down (say -12db from clip) in your mixing application so that all the tracks don't overload the stereo mix. Turning down a track is like backwards normalising, it doesn't undo the rounding errors, it just adds more. Try and map 65536 numbers of loud to 16001 numbers of soft and your kind of missing some loud loud steps by around 4 to 1. Not exacly 4 to 1 either, something nasty like with heaps of numbers after the decimal point.
So you can see that you should never normalise if your goal is to minimise rounding errors and therefore preserve as much of the loudness detail.
Of course a 32bit mixer sounds better than a 16bit one, and many people argue that you don't hear all those errors, but there will be a theoritical limit to how many times you can adjust the gain of a 32bit digital file before you will hear all those errors. My current philosophy is to capture/record/produce all sounds at the highest level and bit depth and minimise gain changes to one or two plugins and mix in the analog world.
You can quote me on any of that, I learn't most of it from 'Audio Technology' anyway. Great Ausie mag!
|