I've asked myself this same question about whether 24 bits is worthwhile for a sample library. For the reasons Charlie stated really well, my opinion is that it isn't. We (Realitone) record at a 24-bit sample rate, but we then normalize all samples, so at that point, my opinion is there's no need for the extra dynamic range that 24-bit offers.
For example, a quiet guitar sample when normalized turns into a very loud guitar sample, so in the mapping editor (or in the scripting), that sample gets volume-reduced by 30 or 40 db so that it will sound right. So the effective dynamic range of that guitar zone is 30 or 40 db plus whatever the dynamic range of 16 bits would be. That's plenty.
So if I were making libraries for myself, where I record 24 bit and then normalize samples individually, I'd make the final samples 16 bit. Partly to save hard drive space (minor issue) and partly for the instrument's RAM footprint (bigger issue). Plus I assume 16-bit is easier on the processor, although I'm not sure about that part.
But ... I'm not making libraries for myself anymore, so not all my decisions can be based on cold hard facts. I'm trying to sell these things, and many potential customers have a lot of preconceived notions. In many people's minds, bigger is always better, whether it's in total gigabytes, or whether it's number of round robins, or whether it's bit depth. So for that handful of sales I might lose to the guys who think our quality isn't up to snuff if we use 16-bits, I stay at 24-bit.
That's all just my opinion, mind you, and by no means have I cracked the mystery of how to run a successful sample library company, so take it for what it's worth. In fact ... having said all that, I'm considering switching (quietly) to 16 bit for an upcoming library where the RAM footprint and processor load will be a major issue.
For example, a quiet guitar sample when normalized turns into a very loud guitar sample, so in the mapping editor (or in the scripting), that sample gets volume-reduced by 30 or 40 db so that it will sound right. So the effective dynamic range of that guitar zone is 30 or 40 db plus whatever the dynamic range of 16 bits would be. That's plenty.
So if I were making libraries for myself, where I record 24 bit and then normalize samples individually, I'd make the final samples 16 bit. Partly to save hard drive space (minor issue) and partly for the instrument's RAM footprint (bigger issue). Plus I assume 16-bit is easier on the processor, although I'm not sure about that part.
But ... I'm not making libraries for myself anymore, so not all my decisions can be based on cold hard facts. I'm trying to sell these things, and many potential customers have a lot of preconceived notions. In many people's minds, bigger is always better, whether it's in total gigabytes, or whether it's number of round robins, or whether it's bit depth. So for that handful of sales I might lose to the guys who think our quality isn't up to snuff if we use 16-bits, I stay at 24-bit.
That's all just my opinion, mind you, and by no means have I cracked the mystery of how to run a successful sample library company, so take it for what it's worth. In fact ... having said all that, I'm considering switching (quietly) to 16 bit for an upcoming library where the RAM footprint and processor load will be a major issue.