Comment by Applejinx

I've been in an internet argument among very serious digital audio experts (such as from Bell Labs) where the consensus reached was this: for properly done audio export as a final stage to be heard by the most critical listeners, and by properly done I mean the output is dithered and not simply truncated and everything else is done properly:

20, possibly 22 bit, and 60 to 80K.

Given that people screw that up by failing to dither to fixed point formats, you could push it to 24 bit, which is a generally supported word length. Since multipliers of common lower sample rates (44.1 and 48) give us 96K, that is also a good 'extra padding' to be certain of never encountering an issue.

I'm with Dan Lavry w.r.t 192K being unnecessary. Done properly, 96K gets everything, including extreme phenomena or artificial sound (for instance, I have a Farfisa organ that's capable of producing reedy thin sounds of extraordinary clarity, from simple electric tone generator circuits). I use 24/96 for my music stream recordings, while also streaming to YouTube at a much lower quality.