🚀 go-pugleaf

If you want the simple answer to this, it's this: Forget about converting to 96k - this is a waste of time. You'll end up with exactly the same sound, only with more samples in it... But if you are going to do any processing at all, though, then you should be working in 32-bit. In any realistic signal, the few samples that you lose to dither are insignificant (since they're at -96dB or thereabouts), and dither will be a considerable improvement on them anyway.

If this is 16-bit material, chances are it wass dithered when created, so you won't be losing a thing. But if you decide to make any level changes in the material you are compiling (for group normalising, for instance), then doing this in 16-bit will lose you far more samples than messing around with dither ever will!

The one thing that I will say about working with 16-bit material in this way is that unless you've done something very strange with it, you shouldn't normally need to re-dither it afterwards, because when you extract the material and convert it to 32-bit, the dither pattern will be converted too. So, assuming that you don't change the levels too much, or apply any really strange EQ (which will alter the dither shape), then simply converting back to 16-bit for your compilation CD without dither should be quite adequate. If you re-dither, you will be adding it on top of the dither that's already there, and there doesn't seem to be much point in that at all, even though you in practice you are unlikely to hear the difference.