In every situation I have heard, 24 bit beats 16 bit.

24 bit recorder to get 16 bits? Well, if you don't record at full level you won't get all the bits and most of us don't record at full level. So, in that situation, you'd be using less than 24 bits, even though you HAD it to use.

Bottom line, having a maximum of 24, even if you don't use them is better than having only 16 and not being able to use all 16.

A LOT depends on the converters. A LOT. When I was testing the Sound Devices 744T, I purposely recorded with peaks (PEAKS) at -30db, just to test the preamps and A/D converters. I then normalized the audio to I could hear it. While there was some hiss, it was low in level. Not may other recorders will allow you to get away with underrecording by that much.

A fictitional test perhaps, but it does say a lot for the converters.

Regards,

Ty Ford

1- My Coles notes version of sampling theory:

When you (down)sample a signal, you have to deal with three issues:
A- Imperfect frequency response. e.g. The highest frequencies aren't as loud as they should be. And you don't necessarily get perfect response up to the Nyquist limit... e.g. 48khz sampling doesn't necessarily give perfect response up to 24khz.
If your sampling rate is sufficiently high, then this isn't really a problem.

B- Aliasing. Weird artifacts if frequencies higher than the Nyquist limit aren't filtered out.

C- Ringing/phase artifacts.
(*Usually not a problem for audio??)

You can get (mostly) rid of two of these problems but not all three at once. So you have to pick which issue you can live with.

To get rid of aliasing, you can either use a bad microphone (one with poor frequency response) and/or you can apply analog and/or digital filtering.
You need to apply some analog filtering before the A-->D converter (and/or assume that the microphone won't produce high frequencies that will alias). Any aliasing that gets into the A-->D converter can't really be gotten rid of.

However, there are some limitations as to analog filtering... there's limitations to what kind of frequency response you can get (and cost considerations). The ideal frequency response would be very good response right up to the Nyquist limit, and then it drops off immediately for frequencies past the Nyquist limit (sometimes called a "brickwall" response).

So your system might implement a mix of analog and digital filtering. With digital filtering, you would oversample the signal... so if you intend on outputting 48khz, you'd sample at some multiple of that rate (e.g. twice / 96khz, four times, etc.). Apply digital filtering to get rid of frequencies above 24khz (the Nyquist limit of the 48khz signal you intended to make), and then downsample the signal to 48khz.
In equipment that does this, the A->D is already sampling at 96khz or some other rate above 48khz. So the higher sampling rate is kind of there already.

2- Anyways I kind of rambled on there. The key point to note is that a reasonable system sampling at 48khz won't give perfect frequency response up to 24khz... you'll get good performance up to some number lower than that. And performance depends on implementation.

Lots more information here:
http://www.wescottdesign.com/article.../sampling.html

3- An oversampled system is not that bad an idea at all since you can avoid both ringing and aliasing artifacts.
Cost-wise, a higher sampling rate does add cost. But to lower the sampling rate, you'd probably want to pick up some ringing artifacts (aliasing won't fly) and you need to apply digital filtering to do that (and digital filtering costs money... so there's a balance there).

(I don't engineer this stuff, so the information above may not be correct. There's a bunch of trade-offs to take into consideration. Like I said... it's the Coles notes version.)

- And to echo AJ's comment...

Glenn,

Having read Jay Rose's books, I'm familiar with some of the arguements for oversampling. It seems that conventional wisdom is: "96K doesn't really help but it definitely can't hurt, so use it if you want to."

Have you heard or found any noticeable benefits to 96K?

Not sure if this has been mentioned already or not, but I just had a discussion with my engineering professor about this. Higher sampling rates reduce the incidence of aliasing when the sound is sampled as a discrete-time sinusoid. While the range is usually from 20Hz to 20kHz, the aliases of the frequencies within this range extend far beyond this range.

There is the theory side of the discussion, and there is the "this is what my ears tell me when listening to reference monitors" side of the discussion.

My comments are purely on the application side, this is what my ears tell me and what other working sound engineers have mentioned to me. 24/48 acquisition* is the sweet spot for many who make their living at this.

*with good mics, good placement, good A/D converters, proper gain structures and a hundred other things that go into making a good recording...