So here is a reason to keep 32-bit floating-point unless you know way better than the average Joe what you and your software are doing.
"Original": A WAV file that I downloaded from an artist's soundcloud site.
Turns out: it was a 32-bit floating-point file.
(I) Converted it to 24-bit FLAC using foobar2000. foobar2000 does not warn that this operation isn't lossless, and indeed it is very far from:
Comparing:
"D:\tmp\Danse Macabre.wv"
"D:\tmp\Danse Macabre_fb2k2flac.flac"
Compared 11871270 samples.
Differences found: 22135695 values, starting at 0:00.000000, peak: 0.2820419 at 2:10.922472, 2ch
Channel difference peaks: 0.2587259 0.2820419
File #1 peaks: 1.2587258 1.2820418
File #2 peaks: 1.0000000 1.0000000
Detected offset as 0 samples.
Difference is 0.2820419, giving you the impression that (nearly) the first two bits are accurate. But no: The WavPack file has a peak of 1.2820419, more than +2 dB, and the conversion brickwalls away this.
(Who said "lossless is lossless"?)
(II) Converted the original using wavpack.exe -b23.9 . That reduces bit-depth to 24 bits. Compare:
Comparing:
"D:\tmp\Danse Macabre.wv"
"D:\tmp\Danse Macabre_wavpack2wavpack.wv"
Compared 11871270 samples.
Differences found: 21259118 values, starting at 0:00.000000, peak: 0.0000002 at 2:10.858481, 1ch
Channel difference peaks: 0.0000002 0.0000002
File #1 peaks: 1.2587258 1.2820418
File #2 peaks: 1.2587256 1.2820418
Detected offset as 0 samples.
The conversion is lossy - and wavpack rightfully warns me - but it is accurate to the bit-depth.
(However it is still a 32-bit floating-point file!)
(III) If I use foobar2000 to convert 32-bit WavPack to 32-bit WavPack with volume normalization through ReplayGain (limit according to peak), I get a file which I can convert to 24-bits at 24 bit accuracy.
But who the hell knows that if they download a .wav, they might have to do (III) before converting to a supposedly "lossless" format, in order to avoid clipping?