24 bit / 96 kHz discussion

  • Thread starter Thread starter jgohman
  • Start date Start date
Igohman: Suggestions on how to evaluate your system without an investment? The gent who runs http://www.pcavtech.com/ appears to use SpectraLAB, a software package that aparently runs under Windows on the machine under test. Check out some of his card test data- it's downright beautiful. I don't know anything about this software at all, other than the fact that it appears to the casual nerd like me to be a very useful tool- and it looks like you can download a freebie test version from the vendor:

http://www.pmgrp.com/prod01.htm

So, if you want to make some serious measurements on your rig, looks like that might be worth a try: "free" is a quality that excuses many faults! That'd probably be more satisfying than trying to hack together differential noise measurements with just your ears- although that is always the final bottom line: can *you* hear it?

I don't have any analog tape machines to bias up any more, but if I did, I'd probably be all over that package muy pronto. It'd be sweet to see the distortion levels and frequency response changes with bias adjustment. Not "strain to hear them", but _see_ them, and know you had it right... Shoot. I knew these dadgum digipooter machines were good for _something_.

Time-domain anomalies: they come in two major flavors. The first is the phase anomaly issue you touch on: as you approach the "wall" frequency, you see more and more phase shift with respect to the unfiltered signal, and sensitive listeners can hear that. After all, these aren't little teeny baby phase shifts: the shift approaches 180deg at the deepest null (since the way you make a null in a filter is to add the input signal back to itself, but 180deg out of phase).

The second is probably best called "time/energy smearing", or just ringing for short. If you put an impulse, or a step, into a filter that has a section with a resonant peak, that section will store the excess HF energy present in the transient and release it over time: PIIiinnggg... Just like a bell, or plucking a ruler that's hanging over the edge of a desk. The excess HF energy doesn't go away, it winds up the resonance and comes back at you later, converted to the frequency of the resonant peak. The higher the Q, the more energy that gets stored up, and the longer the ring as it is gradually released. If you think about it, this is not a phase issue as such: it's really an energy-conversion thing. That's where your ringing on the tops of square waves comes from, since brickwalls by definition require at least a few high-Q filter sections.

You know, it is just downright *bizarre* to be talking about this stuff at all, let alone to be doing it as if I really have any direct knowledge of what's going on today. My knowledge of this field is pretty dated, and I haven't really been active in digital audio at all since about 1986. It was kind of amusing, looking at the comp.dsp FAQ, and seeing them say that the first place to start looking for stuff is Crochiere and Rabiner's "Multirate Digital Signal Processing". I thought "Shit! I _own_ that book!". And there it is, covered with dust on the shelf. Bought it in '83 when it came out, read the parts I needed then, and then put away and didn't touch it again (except to pack it to move, a few times...) until last night. It still has the paper dust cover, even, and it's a good thing: the layer on top was thick enough to bog down a Jeep. I'm a freakin' lousy student.

I've got a lot of catching up to do to understand all the stuff that's going on these days, especially on the high end...
 
Well, sonofabitch. Don't I feel old- and embarassed!

Here I've been typing my butt off for two days describing obsolete technology. I decided to get off of said butt and actually look at what's current, instead of just talking out of said butt about stuff I used to know back when the earth was still cooling.

Point one: when they say they're doing oversampling and sigma-delta conversion these days, they *actually mean it*. Coulda knocked me over with a feather! In the '80s, people used to claim that they did- but the fact was that the hardware was far too expensive and bulky to use in consumer gear, so they just ground the part numbers off the tops of their 14-bit Burr-Brown flash A/Ds, and hoped nobody would really notice that they were padding the 2 LSBs with zeroes....

Point two: this stuff is freakin' ungodly _cheap_ these days. Coulda knocked me over with _half_ a feather on that one, for sure.

I decided to go and actually research the AK5393 and AK4393 converters that are in my D1624. Never bothered to do that before. No kidding: I bought the box because I liked its ease of use and its sound, and didn't pay any attention whatsoever to the claims about its converters. I just flat wasn't thinking that way then, I didn't give two hoots about the guts inside there as long as I could track with it: nerd denial. I'm so used to marketing people lying out the ass that I simply _ignored_ the claims about "128x oversampling", and I wasn't thinking about opening the audio can of worms again- until I found this site, frankly.

The datasheet for those converters is at

http://www.asahi-kasei.co.jp/akm/usa/p_i/html/ak5393.htm

Damn- they're *really doing it*. 128x oversampling means that the basic sampling engine is running at 6.144 MHz: video speed! Guess that the market really has driven things a long way in the intervening 15 years. In retrospect, there's only one word for that: "duh...". The datasheet has a full set of cookbook support circuits, showing the nice, simple first-order antialiasing filter they recommend, with a nice soft knee at 360kHz. Which is all you need when you have a deadband on the order of *6 freakin' megahertz*!

Okay: everything I said about brickwall filters is now relegated to ancient history, looks like. I smite my head with palm of hand, consterned. This is gonna sound funny, but I honestly never would have believed that video-speed DACs and real sigma-delta hardware would have cost-reduced *this far*. According to Dataquest, these things cost $20.25 a pop in 5000s at their introduction in September of 1998, and they do two channels... They're probably half that, now. I had to see it with my own eyes. The bottom line is that I've been living under that rock for way too long- and you guys who are rolling on the floor laughing at me are *fully* justified.

Since they're doing sigma-delta and 128x oversampling, that means that they are getting 9 bits from the oversampling: so the actual raw converter embedded inside the sigma-delta feedback loop (they call it "the modulator") is a 15-bit video-speed flash converter. That would have cost about $1000 in 1984 dollars, *if* you could get it...

Sigma-delta is a nice way of "cheating" in A/D/A conversion. You use a narrower A/D converter and run it faster. You feed back the previous sample to a D/A, look at the previous sample and what's there now (the "sigma" is the sum of the two), and encode the difference (the "delta"). Your narrow converter never has to see a fullscale swing- it only has to see the difference between successive samples. You then pick off this stream of narrow deltas, run it through a digital lowpass filter, and build up your desired resolution with a decimation process (downsampling). Here's a great example of how downsampling can be lossless, for you doubters: you actually _gain_ resolution in this process, by deliberately trading sample rate directly for bit depth.

So you run the sampler at ungodly speeds, and as a result your analog side is basically *completely trivial*. Win-win, if you can get the hardware. And you do lots of DSP to implement the lowpass filter that chops off everything left laying around between your desired output sample rate/2 (FsO/2) to FsI/2. Which is 3.072 MHz- the mind just flat boggles. I'm still having problems with that. And, oh by the way, that filter dumps out the resolution you want. This is 24 bits all the time, in the case of the AK5393- they must dither/truncate down elsewhere inside the box to get to 16 bits, if you ask for it. it doesn't say which, which is all the more reason to keep the full 24-bit resolution until the last minute... Filtering is much tidier to do digitally than in the analog domain, and with that clock rate you can have a lot of DSP ticking along in that little black plastic bug pretty much for free.

Damn, I feel old. No wonder these things have so little in the way of analog artifacting, and sound *so good*. I learned the sigma-delta tricks in college when I took sampled-data systems, but could never justify the cost to design with them- the bean counters would have eaten my liver, with some fava beans and a nice Chianti. And now they are completely trivial, and embedded in $25 sound cards!

Mea maxima culpa. I learned something new today, something about rust sleeping, I think. I'm still scratching my head at how far consumer mixed-signal stuff has come. I shouldn't be, but I am. However, I will correct the folks at Asahi Kasei on one point: the proper name for the technology is "sigma-delta", NOT "delta-sigma", you heathens. You gots to have the sigma before you can create the delta, dammit! At least I learned that much- Professor Oppenheim is probably laughing his ass off at me right now, too...

This is what happens when you don't shut up: you embarass yourself!
 
No,umm no... I really don't believe you need to be embarassed here or is that hear? I believe you first "reviewed" in your eariler posts to bring e-body back REFRESHED and alert for the main dish! The only thing you should be embarassed about is actually apologizing. You are obviously a knowlegable and a humble person whom most of us look to for answers. Hey hold on to that pride and humility - keeps ya on your toes! ( where the hell can you go to and pay for and get the kind of info on this BBS? This is absurd BUT FANTASTIC!)

Perhaps not as verbose as u....... thanx Skippy.
 
Some things to consider.

Sure, lots of people can hear the difference between 16/44.1 and 24/48. A lot fewer can hear the difference between 24/48 and 24/96. Especially if these tracks are not A/B'd. Add to this that most home playback systems can't even reproduce 20 kHz. Also add that the average listener can't even hear above 15 kHz. I have a hard time justifying the hype about about increases above 24/96.

Now I hear rumors of 32/196, do you realize how large that audio file will be? We are definitely reaching a point of diminishing returns. This is all driven by the consumer by the way. If consumers want 32/196, they will get 32/196. The industry has no choice but to respond.

Now, who are the consumers that are driving this trend? Is it the MP3/minidisc/CD masses? No. It's us. The pros and semi-pro. We are driving the demand.

Tom Cram
dbx Senior Technical Support
(801) 568-7530
tcram@dbxpro.com
 
I think its microsoft and intel and chipmakers everywhere that are driving this trend, so that people can keep buying more powerful computers and the latest software.
 
This is sorta like debating the differences between a 2" 16 and a 2" 24 track, and who could hear the difference between the two. What about FM radio, could you ever distinguish between the two on the radio. I suppose these are arguments to be archived back in 76'. Now if I can only achieve "Supertramp" quality from my delta 1010.
 
My point exactly. Pros have always had superior recording/playback systems, it matters to us. Consumers have always had inferior recording/playback systems, it does not matter as much to them. So, WE need to decide when enough is enough. The way we decide is with our pocketbooks. Chip and gear manufacturers are only responding to the consumer, honest. If consumers stop buying the next 700bit/196 gajillion hz unit, we'll stop making them.

dbx makes a digital card for the 500 Silver Series. It only goes to 24/48. We have had exactly two requests for a 24/96 update. We released the 300 series that has built in 24/96 converters because of "market" demand. I haven't talked to a single user that runs it at 24/96...not one. It seems that everybody's talking about 24/96, but hardly anyone is using it.


Tom Cram
dbx Senior Technical Support
(801) 568-7530
tcram@dbxpro.com
 
Tom, I just have to ask how you feel about tube technology. Do you think much of that is is based on a preconception that paople have of attaining a warm sound. To me, the concept of negating a harsh digital sound by sending the signal through a tube first sounds kinda silly. I have enjoyed tubes since the late eighties, back when everyone thought they sounded like crap, but I found that I liked them because they sounded more alive. Anyway, I just found it interesting since the 386 uses tubes and I thought that maybe this is once again, as you pointed, pleasing the demands of the consumer.
 
They are definitely there to please the consumer. Our tube pre-amps outsell our solid state pres by at least 3 to 1. In fact, we discontinued our mid level solid-state pre due to lack of demand. Everybody wants tubes. When a client comes in to my studio and they see a bank of glowing tube pre-amps, they are impressed.

Ironically, our highest-end pre-amp (the 786) is totally solid-state. If I were made of money, all my pres would be 786's. Remember, Neve, SSL, Euphonix, API, etc. are all solid state. The biggest hits in the world are made on the pre's in these boards.

To paraphrase Forrest Gump; Warm is as warm does.
 
Good thread - skippy, how old are you? heterodyning? now that is a term I haven't heard in a while.

"DC to daylight" <- that is a good one, I may have to use that in the class I teach.
 
I'm a good solid fortysomething, well into bifocal terrain and drifting onwards towards shuffleboard...

Feel free to swipe that "DC to daylight" aphorism. I stole it, myself, from a bunch of audiophile types I hung around with in the early '80s. But I've beat that topic to death elsewhere.

I was wondering if I'd get a rise out of anyone with that "heterodyning" comment... While that is a technically incorrect description of the phenomenon, the artifacts sound very much like good old single-conversion AM heterodyning to me. Whee! Let's hear it for "Skippy's Home for Obsolete Technologies"...

After I posted those articles, my wife and I conducted some utterly unscientific tests, using her well-trained voice as the signal source. This seems like a good time to talk about them. I wanted to get a baseline idea of the qualitative differences between sample rates and bit depths using my own rig (which is still rather new to me, since I'm just getting back into this after my Rip Van Nerd time away). Changing rates on the Fostex HDR requires reformatting the disk, and I'm basically lazy, so instead we recorded in mono direct to the Alesis Masterlink, and just tested at the same rates and bit depths that the Fostex can support (16/44.1, 24/44.1, 16/48, 24/48, 16/96, 24/96).

We recorded 6 seperate takes of the same tune at those sample rates and resolutions, and then played them back. What we did was a amateurishly-implemented single-blind experiment (read: scientifically invalid...). I simply didn't tell her which takes were striped at which rates or bit depths (since reconfiguring the Masterlink's converters simply requires pushing two buttons on the faceplate). We could both detect the difference between 16 and 24 bit resolutions, mostly in the reverb tails as you'd expect.

Detecting the differences in sample rates was much more difficult: she was perfectly happy with 44.1, 48, and 96 alike. She's a vocalist, not an engineer, so I'm not overly surprised: with the current crop of oversampling converters that have essentially no analog artifacting in the passband, it really is like picking flyshit out of pepper. I could detect a difference between 44.1 and 48 (or so I tell myself), but 48 to 96 was an absolute wash to my aging ears. Solo vocal music is arguably the wrong sort of signal to use for this sort of testing in any case, because of its low HF energy content: somethign machine-generated would probably be more appropriate. But since that's the bulk of what I'm recording now (not to mention having a willing and attractive signal generator and test subject), there you go.

It was interesting, but ultimately meaningless, in any scientific sense. I ought to tear into the Masterlink to see whose converters they use, now that I've started thinking like an audio nerd again. Look what you guys have done: you've got me curious again... In any case, I feel vastly more justfied in having standardized on 24/48 for my recording. Anything beyond that, I'm personally not able to hear at this point (in my room, with my gear), Therefore, I can also now feel justified to pooh-pooh 96 kHz sampling as a waste of disk space. (;-)

Seriously, though: IMNSHO, going to a 96kHz rate might have made sense to help deal with the filtering issues that are inherent to a non-oversampling converter design- i.e., all that obsolete crap I spouted above. But given that oversampling sigma-delta converters are cheaper than *beer* these days, for gawd's sake, it seems to be much less necessary.

Conclusion: 24/48 is where I'm getting off the bus at this point, but your mileage may certainly vary. Other opinions?
 
On the theory side, I see a definite advantage to 96k (or even, dare I admit it, 192k). Not because of frequency content above 20k, or even for more gentle anti-alias filtering-- but for the better resolution of the content between 10k and 20k. We all want digital EQ's that don't sound like total dog shit, right?

On the practical side, 16/44.1 has been treating me just fine for a while. When I finally do make the move to tracking and mixing at 24 bits, I'm gonna be even happier. I think we're starting to push close to the point of dimishing returns.
 
Another non-scientific test I was playing around with in regards to sample rate.
Download the reaktor demo from http://www.nativeinstruments.com

One of the parameters you can play with is the sample rate.
You can vary it anywhere from 22.05K up to 132K.
There are definitely different sonic qualities at different sample rates - even the high ones - there is a difference between 44 and 66, as well as 66 and 88.

I thought that was kind of interesting.
 
First, thanks for the outstanding posts Skippy. Great stuff to read.

Second, yes it is we as consumers who drive demand for this stuff. But the almighty $$$$$ has a lot to do with it too. If someone came out with a perfect system a whole lot of companies would go out of business. It is the nature of our economy that you must forever buy, buy, buy the latest and greatest that comes out.

As for 32/196, storage solutions are keeping up. The new IBM 100 gig drives should be selling for about $100 by the end of this year. The real bitch is backup solutions which are lagging behind. Soon we will have DVD writers with 18 gig capacity - enough to record one song...
 
Back
Top