44.1 vs 48kHz: Restarting the flamewar...

  • Thread starter Thread starter regebro
  • Start date Start date
regebro

regebro

Insane Genious!
I had a revelation a couple of days ago. I got a theory why recording in 48kHz and the downsampling to 44.1 might actually be better. One reason was mentioned before, that the resolution is higher so that digital effects, primarily reverbs, could work at a higher quality. But that difference is only a couple of percent, so I scratched my head and thought that it didn't feel right.

But here one big difference: The anti-aliasing filters.
When you convert from analog to digital you need to filter away all sounds above half the sampling frequency to avoid aliasing. That frequency is 22.05kHz and 24kHz respectively.
To preserve the sounds at 20kHz, that means that you have to have a very steep filter to block out everything at 22kHz. Steep filters like that are hard to make, and you get phasing problems, and stuff. But if you have 48kHz sampling frequency, you have to block things at 24kHz, so obviously, the filter you use only have to be half as steep to do the same job. So, here is a difference between 44.1 and 48 where the practical difference *doubles* between formats. No skimpy few percents, the filter actually has to be only half as steep in 44.1 and in 48. And of course, if you compare with 96, the filter now only has to have a twelfth of the slope of a 44.1 kHz filter.

In theory you'd like a filter that has 0dB damping at 20kHz and 96dB at 22kHz. But in practice you will cut corners. You'll either start the cutoff before 20kHz, or you will not have full damping at 22kHz. None of these are any real problems, since most people doesn't hear the highest frequencies, and you'll never have a 0dB 22kHz tone anyway.
But with 48kHz, you don't have to cut corners as much.
I suspect that most sound cards would use the same filter in both cases, actually, either raising the cutoff-frequency by 2kHz when doing 48 kHz, or have a cutoff of maybe 30dB at 22kHz and 60dB at 24kHz, which with high frequency sounds could cause some minute aliasing.

The slight loss in HF content, or the minute nearly inaudible aliasing might be enough to make the sound be percievably worse, especielly when processed.

But as I said, this is only theory. What do you think? Sounds reasonable?
 
That's a good point you have. I think there are more aspects. Should check the books from my course last year again...

But let's not forget the marketing aspect. For some obscure reaon, DAT was designed at 48 kHz, and CD's at 44.1 kHz. The best quality you can get is when using DAT. So this is what people used.

Another thing is: DAT never really made the consumer market, so most equipement was developped aiming at the professional market, studios etc... So they HAD to go for the best quality they could get. So recording at with this material and eliminating the lower quality stuff had to give better results.

I don't know how much of this is still true today... Quality gets cheaper and cheaper...
 
well regebro,

i'm happy to say your wrong! skippy posted the same idea but later came back on it cuz of oversampling it wasn't true anymore. don't know the name of the thread but skippy writes excellent posts anyway!

(ps. don't argue with me i wouldn't know... just copying here!)
 
Hhm, yeah. But does most soundcard manufacturers really use oversampling? Creative makes no notes of this anywere, for example. The high-end cards does, of course, but then they usually support 96 kHz too, which makes the choice between 44.1 and 48 rather moot. :)

But just about now, something else completely just struck me...
1. I know we had problems with syncing audio and midi when we first tried using HD recording. When this syncing is done, HOW is it done? I assume that its doesn't try to stretch the sound, but syncs the midi to the exact sampling frequency of the sound card, right?

2. The soundcard really has to sample on a multiple of 44.1 kHz when is samples. A card that oversamples say, 24 times at 48 kHz but uses 1.152 also MHz as the sampling frequency for 44.1 and then interpolates to 44.1 would be much easier to when constructing, but then you still get interpolation. In that case, choosing 48 would make sense.
 
well your right about the sb thing.
But all cards that (at least) pretend they'r for serious audio use oversampling.
i'm not sure bout question 1 as i'm not that long in the wonderful world of HDrecording ;) but i don't think their is something like "stretching" audio to fit midi commands. the other way around seems MUCH more atractive. I guess you'd better ask skippy. That dude knows sooooo much.

in my eyes the main advantage of using 48 instead of 44.1 is the better resolution for the effects as you previously mentioned. And a big part is IMHO the manufacturers who want to create a need.

but uhm, as i said; i don't know shit. ask skippy.

greetz guhlenn:)
 
Oh, yeah. I really stepped on my pudendum on this one, talking about ancient history before I went off and did the legwork to find out what was actually the current state-of-the-art. If you'd like to see the whole hilarious and regrettable episode, here's the pointer: https://homerecording.com/bbs/showthread.php?threadid=21183

My original assertion was that straight sampling requires a brickwall filter, and that the additional deadband provided by samping at 48kHz allows a more relaxed filter design. This is certainly true, and has been since the dawn of digital audio (which is, regrettably, where I got off the bus...). All the designs I was involved in in the early '80s did straight sampling, period. That's really all that was available.

So that's what I posted. And then I got that funny itching feeling at the nape of my neck: the one that unmistakeably tells you that your gut knows you screwed up, even if your brain is at Costco or something. I went off and did some real research on what is current out there, and found that oversampling converters have been cost reduced to the point that *excellent* 24/96 converters can be had for under $10 per channel in quantity. At that point, the analog electronics to do a decent brickwall cost more than the current crop of fully-integrated monolithic high-res converters. Which makes it a no-brainer, if you're a design engineer: use the oversampling stuff, because it's now cheap.

Argh. In retrospect, it should have been completely obvious, even to me. Sometimes it takes 20 years for ideas to make it through my skull by osmosis... But here it is: if the advertised resolution of a converter is 20 bits or greater, it uses oversampling, guaranteed. And probably the vast majority of 16-bit converters out there these days do as well.

The reason is simple: straight sampling A/D converters with more than 16 bits of *usable, linear* range are very hard to make, and are very expensive. Pop open a Newark catalog and have a look: all the regular straight-sampling linear converters stop at 16 bits, just like they always have (and even so, the linearities barely make it to +- 1.5LSBs...), and they still cost many tens of bucks. But all over the place you see oversampling sigma-delta converters that do 20, 22, and 24 bit resolutions with bandwidths of 500Khz and up.

The bottom line is that the pro audio industry took a hard left turn after I quit paying any attention to it. You can't make a cost-effective monolithic D/A with a resolution over 16 bits with current semiconductor processing technology. However, the important thing is that _you don't have to_: you can make one hell of a *fast* 14-bitter, and run it inside a sigma-delta loop that oversamples by oh maybe 128x, and then use a digital decimation filter to trade off output sample rate for output bit depth. Viola', a 24-bit converter. Now that all that functionality is all packed into one of those cheap little black plastic bugs with all them legs, it's a true no-brainer to use.

This is the funny part. I design microprocessors and DSP engines for a living: that's what I've been doing for the last 20 years since I bailed on the pro-audio/pro-MI market. I have done some processor designs that run at well over 1Ghz and are pretty cheap, given what they do. I live with the nasty realities of semiconductor process variations every day of the week. And even so, it never *once* occurred to me before I posted that thread that the consumer audio market had come along at the same or faster pace, until after I hit "post it". 24-bit linear? What the *hell* was I thinking. You can't make it that accurate, even with laser-trimming: but you can make it hell-for-fast.

There is one word for that realization: "Duh!" Mea maxima dazed old fart... Open mouth, insert foot up to the navel. (;-)

Sync is pretty easy: as fast as processors are nowadays, that sort of realtime work is much easier than it used to be. You would simply install a lookup table for internal delays, and move the timing of the transmitted MIDI event to line up with the audio stream (which you cannot muck with the timing of without causing some really gnarly artifacting).

And oversampling really makes picking different output sample rates trivial: you actually sample at a large multiple of the design-center output rate, and then you can just change the filter coefficients for the decimation filter to pick your final output sample rate. The 128x AK converters I have in both my D1624 and my Masterlink sample at 6.144 MHz for both 48kHz and 96kHz output sample rates, and just change the filter to select the final downsample. For 44.1kHz, they shift the input clock frequency to 5.645Mhz.

And Creative doesn't mention it because they don't want to scare off the techno-peasants, I'm sure. Sooner or later that jargon will get mainstreamed. But right now it's not necessary for Joe and Mary Sixpack to know that the audio for Quake is produced for their listening pleasure by an oversampling process (as opposed to perhaps a bunch of tiny gnomes with tiny drums and stuff....). (;-) Note the smiley, please, I really am being sarcastic here. But oversampling isn't a useful marketing term for Joe and Mary Sixpack- only for you and me, and only for me if I can get my femur away from my uvula so that I can see the freakin' *screen*.

[Edited by skippy on 02-19-2001 at 09:10]
 
it always is a pleasure to read your in-depth yet clear posts skippy! thanks!

greetz guhlenn
 
Back
Top