Topic: Help me understand 32-bit float

[acidjack]

When I edit in Audition (CS 5.5, Mac) I notice that the multitrack file defaults to 32-bit float.  I think this is what most editors do.   Then, when I export back out the mixed file, it has to dither back to 24bit. 

I assume this is the best way to do it, but what would happen if I just made the multitrack file 24bit also, thereby not having to dither it?

[stevetoney]

Gosh I never thought about it because I just assumed that the multi-track default was the wrong setting for my mastering needs.  Since my source files are always the 24 bit IEEE whatever, I changed the default of the multitrack mix-down files to match so that the dither wasn't needed.  I have no idea though if my initial assumption is right or not, so now you've got me curious to the answer to your question too.

[Gutbucket]

Let it do it's 32 bit thing internally.  The editor should take care of truncating and dithering down to 24 or 16 bits or whatever you specify for your exported files.  32 bit float allows sufficient internal processing headroom for the calculation math involved in changing levels, summing channels, and all other calculations done on the files.  Once those calculations are done, the 32bit >  dither and truncate > 24bit output step is harmless and better than forcing the editor to estimate and round off the math to work within just 24 bits.

Think of it as analogous to the well understood advantage of recording 24bit files even though the release format will be 16 bits.  In that case 16 bits is fully capable of containing the full dynamic range of the music, but the additional headroom is welcome and helpful in relaxing the requirement to get levels set perfectly.  In the case of the editor, if it is working with 24 bit files to start with, it needs more than 24 bits of calculation space to fully accommodate the calculations.  Once the calculations are done, the data can easily fit in 24 or 16 bits.  It's about transparency of processing.

Similarly on the sample rate side of things, many plugins have the option to oversample the incoming data before doing their processing (allowing for less step filtering which is likely to be more transparent and artifact-free) they then sample rate convert back down again to the original sample rate of the file at their output.

Most editors have the option to save a 32bit version, which might make sense if storing an intermediate working version of files that are still being edited, but makes no sense as a final output format as it would be total overkill.

[page]

yeah (i would have said the same thing).

[acidjack]

Thanks guys - very helpful.  Glad I asked.

[stevetoney]

Thanks guys - very helpful.  Glad I asked.

X2

[dallman]

Interesting, but it makes me wonder why this would not be the case in Soundforge. I do all my editing in 24 bit (24/48 in my case) and then dither to 16 bit (16/44.1) usually for the final product although I keep the 24 bit file too.

[Gutbucket]

I don't use Soundforge so I can't say for certain, but most editors work in 'longer word-lengths' behinds the scenes and it may not be apparent that they are doing so unless you dig.  In otherwords your input files may be 24 bit and report as such when you look at their properties within the program, and your output files maybe as well, but the program is still likely to be using longer words internally when doing it's processing.  Older more primitive versions of editors did not, but doing so has been pretty standard for a while now I'm pretty sure.  Can any software experts confirm and put a aproximate date on that milestone?

[SmokinJoe]

Do you remember a little high school math/science.... 6.02x10^23 and all that?  A standard 32bit float as represented in computer memory has 23bits of abscissa (the 6.02 part goes out to a bunch of decimal places) and the rest is exponent (and +/- sign bits).  So they take your 24bit number and convert it to a 23bit number, plus the x10^y. If you really want to know the details, google knows all.

Why?  Because modern computers have floating point processors which makes them really good at doing 32bit float math.  They can do that kind of math a lot quicker than integer math, and most people want computers to run fast.  In Audacity it's pretty evident.... if you force the format to 24bit instead of 32bit float, operations like "amplify 3db" will take significantly longer, maybe even twice as long.

When it comes to storage, you only have 23 or 24bits of information, so store it that way, storing it as 32bit format will only take up more space, by a factor of 1.33 to 1.

I really really hope we don't have a bunch threads about dithering between 23 bit and 24bit, because it's just plain silly. Nobody can hear that difference.

[H₂O]

Everything should be mastered in "future proof" 1-bit..... HA