What happens if you go from 32 bit fp to 24 bit fixed
Floating point calculations allow for more precision in lower values. 32 bit floating point calculates sample points closer to zero with 24 bits. In fixed point 24 bit you lose 1 bit of precision for every 6dB you go below 0dBfs.
In this video I import a normalised snare sample and measure its bit depth - 16 bit - with the plugins Bitter and Good Dither.
After turning on Space Designer and Compressor, the perceived loudness goes up. Bitter and Good Dither show that the audio signal is no longer 16 bits deep, it's 24 bits deep and its values slide across the scale due to floating point calculation. The "cloud" of bits descends in the tail of the snare sound.
When I bounce in place, which turns the signal into 24 bit fixed point, Bitter and Good Dither show that only the peak is represented by 24 bits in the calculation and after that more and more of the bits available have value "zero" which makes them useless for precision.
The values in the signal that are too low for 24 bit fixed point get rounded off (truncated) which causes distortion patterns that are too low to distinguish at first.
Further processing of the distorted math results in a buildup of mistakes that gets really annoying, unless you apply unshaped dither before you truncate from 32bit fp to 24bit. Mastering engineers like Bob Katz, Ian Shepherd and Bob Olhsson have posted about this on the web.
When I "freeze" the file, the signal does not lose precision.