FYI: Reasons for digitizing at 96kHz or above

Some Archives work intruding -- I do digitization of analog tape recordings for both preservation and use copies. And in the research I've done, few people talk about why sampling frequency needs to go up, and up, and up -- I hear a lot of "96kHz is marketing -- you can't hear sounds up there anyways." And I'm not sure that 192kHz isn't marketing, but having the equipment now, I can tell you why high-frequency digitization is an absolute necessity.

First, a bit about Nyquist-Shannon theory. If the signal you want to faithfully reproduce is at frequency X, you need a sampling frequency of at least 2X to reproduce it. In practice, you actually need headroom over that, especially for dither. So, in your Red Book-standard CD, the audio is 16-bit, 44.1kHz. Now, that specification is the way it is because of older standards of mastering audio recordings, but what it means is that in that 44,100 Hz of headroom, you can reproduce sounds up to 22,050 Hz -- in practice, you only really get good stuff to about 20kHz, but we usually talk about that as the limit of audible frequencies anyways. So you can see that 48kHz (DVD standard) gets you reproduction to just under 24kHz, and 96kHz gets you reproduction to just under 48kHz.

So why do you need all of that range, when the ear picks up on only the stuff between 20-20,000 Hz?

Here's what I've found: basically, even if you can't hear it, it's still there in the analog signal. It doesn't matter, to start with, what frequencies you're dealing with as desirable -- speech voice recordings rarely have anything useful in them above 12.5kHz, for example. The noise runs through a much broader range of frequencies. 44.1 and even 48 don't accurately reproduce anything higher than 24 kHz, and everything above that has artifacts. You wind up with the frequency equivalent of amplitude clipping. It can't be removed, even with hard low-pass filters, and it can't be repaired. Often, we're talking about noise in the HF bands that isn't at audible amplitudes, say around -88dB. And nobody in the analog world would ever worry about that! It had as well not exist, because every component in the chain contributes more noise than that. But digital sampling doesn't care about audibility, it cares about mathematics.

So the moral of the story is, take your initial digitization at high frequency, 96kHz or better, because you need to faithfully reproduce the full noise spectrum to do anything of useful quality with the desired signal. After that, if you're cleaning out the noise, make sure you roll off anything above your Nyquist limit before you downsample, and everything will be great!


Popular Posts