The catch is you can't shape the noise without making it. That is, you can't take a 24-bit audio file and turn it into a 24-bit file with the noise rearranged with noise shaping. You could turn it into a 16-bit file with noise shaping, but it will be noisier than the 24-bit original.I think what DonH56 is saying, is yes upsampling can reduce quantization noise. If you have noise above quantization noise in the original file, then yes it will still be there. However, the part due to quantization will be less if upsampled properly. Maybe it is covered by other noise, like a digital file of an analog tape, but the quantization caused noise will be lowered. Even without upsampling, you can save digital files with different dither levels like shaped dither which can alter noise floors caused by the quantization errors.
With shaped dither you can create and recover signals below -96 db FS with 16 bits. How far below will depend upon how the dither is done or how much upsampling you do.
That is, noise shaping is done by processing the error, redistributing it spectrally. But if you're given a 24-bit file and the goal is to make a 24-bit file, you don't know what the error is for each sample—it's all just signal at that point. But if you're goal is to make a 16-bit file (or even a 23-bit file), the error for each sample is the difference between the 24-bit and the target quantization. You can shape that.
Similar thing. Upsampling doesn't spread the noise, because the new samples have similar error to the original samples—because each sample has the same average error (± half lsb). Twice as many bits, the same error bounds, spread over twice the bandwidth—you're back to the same place. If you do something else, like quantize to a lower bit level, that noise would be spread out. But if you quantize to a lower bit level, you can do error shaping and there's not much point in upsampling.Upsampling cannot add more to the music, but can reduce the noise, if you keep the signal bandwidth the same when you upsample. That increases the SNR (signal to noise ratio).
Anyway, I mainly wanted to say that none of this matters at 24-bit. Whatever you recorded has far more noise. It does matter for 16-bit, where quantization noise can be the largest error component.
I want to say this because I'm certain a few people will read these things and conclude that they need to listen to higher sample rates for a better noise level. Just listen to 24-bit and there is nothing further to discuss.
NB—I'm not saying everyone needs to listen to 24-bit. I'm saying if you're concerned about things like dither and noise shaping and noise spread over a larger bandwidth, then going 24-bit is only 50% more data than 16-bit and completely eliminates having to fret about the math—you're guaranteed to be better than the electronics and your ears. Quite a bargain