Hi Brian,
this is definitely not simple matters. I am not sure I really understand it and definitely certain that I cannot really explain it. I´ll do my best though. The matter is not quite intuitive. Some parts of it seems to be quite against common sense.
First, let us assume that we are measuring a sound source going between -1 and +1. It could be Volts, the units is not the important thing here.
At every moment in time we want to measure it as exactly as possible. Depending on the number of bits we have in the AD the resolution we have is (figures are rounded):
16 bit : 0.00003
20 bit : 0.000002
24 bit : 0.0000001
What this means is that we cannot "see" a difference smaller than that.
Now, add noise into the figure. If we measure a Sm (measured signal) it is actually composed of S + N (the true Signal + Noise). The noise is a (more or less) random figure, it could be either plus or minus. Let us assume that the N part goes between +0.0001 and -0.0001. What this tells us is that, in general, the smallest bits we are measuring are totally swamped by noise. Regardless if we use 16 or 24 bits in that example, the last few bits are still only noise.
Let us go back to the figure S/N or signal to noise ratio. We have to do a mathematical operation here to go from number of bits to an S/N value. But as large mathematicians has already done that we can rely on them. It turns out that one bit represents about 6 dB. So if we have 16 bits we can have a maximum of 6 x 16 = 96dB of S / N. In effect this is a different way to say that the smallest measurement we can make is 0.00003.
As the noise is composed of many parts, we might want to look at them. First there is noise in the room -- a very quiet room may have 30 to 40 dB of noise. And if the maximum sound volume we want to have is, say 120 dB, then the room has something like 80dB S/N or about 13 bits. (And it is not quite as simple).
Next comes the mic. Smaller mics generally has more noise than larger (this is partly due the laws of nature, maybe to be talked about at a later time). Good small diameter condensers turn in at about 20dB equivalent noise level, really small electret mics may have much more. Add to that mic pre amps that can only amplify and add noise, never remove it.
The problem with noise signals is that they add together (adding dB is not a simple adding operation though in this case, as we are adding different noises). In the end we might find out that the 40dB Noise of the room has gone up to, say 50dB, but we still have a maximum of 120dB. So we are now down to 70dB of S/N, ie around 12 bits.
Now, nothing is as simple as the above examples. The ear will actually hear things a bit below the noise floor as expressed in bits. We will probably waste one or two bits at the top as we keep signal levels conservative to insure that we never clip the signal (which sounds horrible). But all in all, for typical amplified music, 16 bits is definitely OK if we are sure that our chain gives us that.
For classical music, we want a little more. 20 bits is about enough for most all classical music, an about what we can get from top-level stuff today.
Personally I preferr 20 bits (well, ok then, my machine calls it 24 bits, but 20 is about what I get). The simple reason is that I want to be safe and lazy. Safe against going up into the red territory where things start clipping. Lazy enough to simply set a recording level and then let it run. I generally aim at -12dB for the peaks, they tend to go a little over that from some reason. This will "eat" about 3 bits, but there is still 17 left to go, plenty really.
Could there be other reasons to record with more than 16 bits? Probably, for ease of postprocessing or perhaps simply because the box we record might sound different at different bit depth.
Sorry, not quite clear presentation, done my best though.
Gunnar