Topic: 24-Bit / 48kHz or 96kHz

[justink]

right, so the higher the khz (ie.  96 vs. 48), the better, right?  the more samples you have, the closer you are to the actual analog signal/wav, right? 

Not really.  You can't hear above ~ 20 kHz, which is perfectly reproduced (no information lost) at a 40 kHz sampling rate.  At 96 kHz, you aren't reproducing the wave at 20 kHz any better and the additional samples go to reproducing ultrasonic frequencies.  If your mics are recording them in the first place...

The Lavry paper I linked earlier explains all this in considerable detail.

i guess i've just had the whole "theory" of it wrong to begin with... good info, thanks.

[aaronji]

You can perfectly reproduce a wave form up to a given frequency by sampling at twice that frequency.

And importantly that signal needs to be bandwidth-limited, otherwise you will get aliasing.  The required filter takes up a bit of bandwidth, and the more DSP resources you have for the calculation gives you better accuracy, lower latency, or both.  So those are the physical constraints that might cause one to select a sample rate greater than twice bandwidth.

My (admittedly limited) understanding of this is that you don't have to go too far above Nyquist to allow room for that filtering with modern ADCs?  Any idea how much?

[SmokinJoe]

24-Bit over 16-Bit gives you better headroom and if you need to boost levels, you'll boost less hiss.

The above is a common myth that keeps getting repeated, but that doesn't make it fact.  Back in the day when people ran analog tape that was true, because you did lose headroom and get hiss.  In early digital media (DAT etc) they maintained a similar mind set.  With digital wav files if you want to run at -6 or -12, go ahead, it's easy to click "amplify" on the computer afterward.  Because it's easy people said "dude, stop stressing about clipping, just run it at -6 or -12 and bring it up in post."  That's not really because of 24bit, that's just coincidental timing.

A/D converters aren't magic.  Let's imagine you are a kid in 3rd grade with a curve on graph paper.  You interpolate your curve into little XY points.  That's all an A/D converter does.  The real magic is all in the analog stuff that makes the graph in the first place.



The top and bottom extremes are the limits of your A/D converter... i.e. 0db = where clipping starts.   If you run 6 db below that, you have 6db of headroom, which is helpful to avoid clipping on the occasional drum whack. How much hiss you have is a function of how noisy your analog gear is before the A/D.  Headroom and hiss are unrelated to A/D converter, sampling rate, or sample depth, they are in the analog domain.

Using 4, 8, 16, or 24bit is just like using much finer graph paper along the vertical scale.  Sample rate is how coarse your graph paper is along the horizontal axis.  Now it's absolutely true that higher resolution will allow you to record the numbers in smaller increments and then if you plot the points and redraw the curve it will be more accurate, at least along the vertical axis.  At some point you ask "how much is enough, and when does it become overkill?" Have you listened to some of those 8bit PCM GD tapes from years ago?  Personally, I think they sound pretty damn good!!!  Going from 8 bit to 16 bit is 256 x smaller increments, and that's significant.  Going from 16bit to 24bit is another 256 times, and at some point it becomes overkill.  It's like a 4megapixel camera versus 1000megapixels.   More is better, but at some point enough is enough. So 16bit versus 24bit has merit, but how much merit is anyone's opinion.

[page]

44.1k became a standard because the guys who created the standard CD wanted a particular symphony to fit on 1 disk and that made it fit.  But at a technical level I'm sure they agreed it was in fact "good enough".  48k became a standard to go with video tape.

urban legend from what I gather, the size was based on the cassette as a basis and the sample and bitrates came as technical hold-overs from prior technology that they built upon.

The AES history paper on it is a neat read (available here):
http://www.exp-math.uni-essen.de/~immink/pdf/cdstory.htm

but yes, the rest of the post is well done. (I thought the old Sony ADCs used in PCM were 12bit or 14bit? I don't know)

[Gutbucket]

Using 4, 8, 16, or 24bit is just like using much finer graph paper along the vertical scale.  Sample rate is how coarse your graph paper is along the horizontal axis.  Now it's absolutely true that higher resolution will allow you to record the numbers in smaller increments and then if you plot the points and redraw the curve it will be more accurate, at least along the vertical axis.  At some point you ask "how much is enough, and when does it become overkill?" Have you listened to some of those 8bit PCM GD tapes from years ago?  Personally, I think they sound pretty damn good!!!  Going from 8 bit to 16 bit is 256 x smaller increments, and that's significant.  Going from 16bit to 24bit is another 256 times, and at some point it becomes overkill.  It's like a 4megapixel camera versus 1000megapixels.   More is better, but at some point enough is enough. So 16bit versus 24bit has merit, but how much merit is anyone's opinion.

This is incorrect as I understand it. PCM encoding to a higher bit depth provides a larger range of possible signal levels which can be stored.  It does not increase the resolution of information within the range which is covered by a lower bit depth, it expands the upper limit of that range.  The fact that the amplitude variations of most music range around 30dB or so for the most part (some more, some less, exceptions for fade outs and very dynamic stuff like some classical music) and that the proper use of dither can allow for the perception of signal well below the noise floor, partly explains how music can still sound very good through limited bit depth formats.  What was the early Denon digital recording system? Something like 12 or 13 bits? That's approximately equivalent the range of level available with analog LPs.

[SmokinJoe]

This is incorrect as I understand it. PCM encoding to a higher bit depth provides a larger range of possible signal levels which can be stored.  It does not increase the resolution of information within the range which is covered by a lower bit depth, it expands the upper limit of that range.  The fact that the amplitude variations of most music range around 30dB or so for the most part (some more, some less, exceptions for fade outs and very dynamic stuff like some classical music) and that the proper use of dither can allow for the perception of signal well below the noise floor, partly explains how music can still sound very good through limited bit depth formats.  What was the early Denon digital recording system? Something like 12 or 13 bits? That's approximately equivalent the range of level available with analog LPs.

I think it's matter of perspective.  As far as the actual numbers go, an 8bit A/D will divide "full range" into possible numbers of 0 to 2^8 -1 = 0 to 255.  12 bit is 0 - 4095.  16 bit is 0 - 65535, and 24bit is 0 - 16,777,215.  So whatever "full scale" is (+/-  X volts) get's divided down into that number of pieces.  you might say "the number range goes higher".  I say we divide into smaller increments of 1/x... same thing, different words.  I once heard it described like so... think of 24 bit as the 16 bit range 0 - 65535, and then have a fractional x/256th's appended to that.  For some people that might be easier to grasp... it's really all the same.

Dithering is a whole 'nuther issue.  If you have data available at a high bit depth and save it to a lower bit depth, it's probably a good idea.  That is a big selling point to audiophiles, but I think it's only slightly more important than "burning in your cables." For years a lot of us ran UA-5's > JB3/H120/D8 which truncated with no dithering.  It sounded fine if the analog part sounded fine.  If you record at 24bit, edit at 24 bit, and then save to 16bit the software dithers... I mean it's free in software, go ahead and use it.   I've exported samples out of Audacity with and without dither, alternate the samples, and if I listen with my eyes closed I can't hear it. When you are going to less than 16 bit it probably makes a bigger difference.

Yes, I realize this is all signed integers instead of unsigned int, but why confuse the matter.

[Gutbucket]

Yep two seperate issues.

I think one important point on dithering and real world use is that most anything we record will have analog noise that is significantly higher in level than the noise floor of the ADC, so when we record that analog noise effectively acts as very loud dither taking care of any quantization noise which would be much lower level anyway.  But all modern ADCs dither as do the softwares we use to manupulate things once recorded so that's really a dead issue and I think we agreed there anway.

But on the bit depth thing you are right in one sense that it's a matter of perspective and could be set up either way, but the reality is that 'full range' is set by the engineering standards of the medium and we don't get to change it at will.  Theoretically we could make an 8 bit system with a 100dB dynamic range which has very coarse loudness changes, or 32bit sytem with a total range of 50dB and way more incremental levels than necessary, but neither are real world options. PCM gear all uses the same 'standard' of each bit representing, what is it? something close to 6dB of additional range.   So with 16 bits we get something like 96dB of range between noise and full scale, and with 24 bits we get 144dB of range..  at least in the digital realm.  No analog equipment has that much range so the best we can get in reality with great studio converters is closer to 125dB or something, which would fit in about 21bits.  But 24 is an even multiple of 8 so it makes for easier computer manipulation of the files and isn't that wasteful with a few bits of noise at the bottom in the most optimistic case, or quite a few bits of noise down there with most anything we record around here.

[Gutbucket]

Just thought of something that might help make that point.  Jon please correct me if this is wrong.

We've all had files with incorrect headers play at the wrong rate before.  Say a recording made with a sample rate of 48kHz played back at 44.1 so that it sounded slowed and pitch flat. When converting from one rate to another we need to do some math to divide one into the other, it's not like changing the bit depth of the file where we can just truncate the extra digits (and it still works with or without dither  )

If the change of signal amplitude value stored by each bit wasn't the same with 16 and 24 bit files, then we'd need algorithms to bit rate convert from one depth to the other.  Simply truncating the least significant bits wouldn't work.  If the amplitude value stored by each bit was different, we'd get the dynamic equivalent of playing back the file at the wrong sample rate- instead of pitch and speed change we'd get dynamic range expansion or compression.

Basically the non-linear transfer funtion of dynamic range Jon mentions.. I think!?  Really reaching the limits of my knowlege here.

[F.O.Bean]

I record 24/48 99% of the time. When I'm recording plays, orchestras, church recitals, acoustic stuff, etc, THEN I record 24/96

But IMO 24/48 is sufficient enough for recording shitty PA Systems

[taperdave]

ok, my question is this: If battery life and card capacity are not an issue and your DAW can handle the file sizes, etc. Is there any harm in recording 24/96 ?
I use Wavelab 6/Waves Mercury for mastering and dither etc.
I thought after reading through 5 pages of reasoned discourse that SmokinJoe had handed down the truth from on high  , I mean, he had a graph for the love of strange medicine. 
But alas even that soon was shrouded in uncertainty   
So is there a reason not to go 24/96 all the time, if capacity isn't an issue?
Dave
readin a lot, synapses just fire too slowly these days

[mosquito]

...
So is there a reason not to go 24/96 all the time...?
...

(If you like, just read the bold parts.)

So, to rephrase, if there were no issues of cost or time or power supply or data storage or output format conversion etc., and the only concern is getting the best possible recording,
is it better to use 24/96 than 24/48?

Maybe?  Honestly, I *don't* have a full understanding of all of the nitty-gritty details, but in the Lavry paper, http://www.lavryengineering.com/documents/Sampling_Theory.pdf , he's saying that sample rates like 192 kHz do not provide any better audio information and can increase errors. 
Lavry doesn't speak directly on 96 kHz.

My suspicion is that it ultimately depends on what any particular ADC can do.  If it's an inexpensive model like mine are, they're probably inferior and will probably create sampling errors at 96 kHz much like Lavry speaks about happening at 192 kHz in other devices. 
I suspect my low-end DR-100 will create more errors at 96 kHz than it will create a better representation of the analogue signal.  I bet the only way to really know is to test my particular unit.

[SmokinJoe]

But on the bit depth thing you are right in one sense that it's a matter of perspective and could be set up either way, but the reality is that 'full range' is set by the engineering standards of the medium and we don't get to change it at will.  Theoretically we could make an 8 bit system with a 100dB dynamic range which has very coarse loudness changes, or 32bit sytem with a total range of 50dB and way more incremental levels than necessary, but neither are real world options.

I think we can choose the "full range" of the A/D for our application, or amplify/attenuate the signal to match the A/D.  On the industrial side, I choose a 0-10v A/D to read a pressure transducer where 0-10v=0-100psi, but I use a different A/D to read millivolts and thermocouples.  The pressure transducer probably has a strain gauge on a diaphragm instead of a capacitance based capsule, but that tiny signal is amplified to match my A/D.  I could plug a ribbon mic directly into the back of my HD24... but it's a poor application because I know the mics output is small compared to my A/D's input range of +22dbu.  Most of us choose a packaged audio system, like a V3, or an R44 recorder, and we let the manufacturer make those choices for us.  I have no idea what it's internal A/D range is, and I don't really care, I adjust that analog section to match the signal to the A/D's range.  It's all just bits between +/- full scale.

Going back to my previous statements (2 pages ago) I'm really isolating just the A/D and disregarding any analog portion of the circuit.  Dynamic range and frequency response are attributes of the analog signal, either before the A/D or the digital equivalent after conversion.  The A/D just plots those XY points as accurately as it can at any given moment in time.  If it's a positive voltage you get positive numbers, negative voltage = negative numbers.  If the signal has low amplitude, you end up with small numbers, and you sample that rapidly.  What someone said a couple of pages ago was that 24bit by itself gives you more headroom, I say no... headroom belongs to the analog realm... I adjust the analog to match the A/D, headroom helps avoid clipping regardless of the A/D range.  If you want to mess with the signal do it at the analog stage before the A/D, or in a computer after the A/D.  But the A/D stage should be as accurate as possible.

I need to amend my statement about "hiss", where I assumed a real world practicality to "tapers" (99% of the people on this board), but I shouldn't have generalized.  In any kind of live setting there is considerable noise in the room... whether it's chatty drunks, or people sitting quietly listening... there is no such thing as "dead quiet".  People are breathing and other stuff... and that is generally louder than the self noise of my mics or preamp...  I never come close to 96db dynamic range, so running 16bit with 6 or 12 db of headroom is fine.  That applies to 99% of us, but not all.  In a studio with good analog gear, the quantization noise from 16 bit could get you, and then dithering is more important.  My primary beef is with the posts who make the 16bit tapers feel like second class citizens... they keep regurgitating the same rhetoric fed to them without thinking or understanding, and it's not relevant to most of the people on this board.  For most of us, a 16bit A/D would not be the weakest link in the chain.

If the change of signal amplitude value stored by each bit wasn't the same with 16 and 24 bit files, then we'd need algorithms to bit rate convert from one depth to the other.  Simply truncating the least significant bits wouldn't work.  If the amplitude value stored by each bit was different, we'd get the dynamic equivalent of playing back the file at the wrong sample rate- instead of pitch and speed change we'd get dynamic range expansion or compression.

Again it's all just percentage from +/- max, like sine wave values range from -1 to +1.  The algorithms to convert from one bit depth to another are so routine we just take them for granted.  For 16 > 24 bit, you simply pad with nulls.  That's an algorithm.  24bit > 16 you can truncate or apply a wide variety of dithering algorithms.  Stored in a wav file header is a block which tells you if you are reading 16bit or 24bit data... get that header wrong, and its scrambled digi-noise.  As I recall the 24bit RIFF file format is something like 16bit data for Left, 16bit data for right, than the 8 bits of LSD for right, and left.  Which is where my other analogy of 16bit data with 1/256 fractions comes in.  I'm familiar with it because I wrote code which rips through files to create what I hoped would be a better declipper/mousetrap (trying to save a particular recording, which didn't help much).

[hi and lo]

My suspicion is that it ultimately depends on what any particular ADC can do.  If it's an inexpensive model like mine are, they're probably inferior and will probably create sampling errors at 96 kHz much like Lavry speaks about happening at 192 kHz in other devices. 
I suspect my low-end DR-100 will create more errors at 96 kHz than it will create a better representation of the analogue signal.  I bet the only way to really know is to test my particular unit.

I think that's probably an incorrect assumption. The argument made against 192kHz, as I understand it, is not economical in nature or related to quality control, but rather that a sampling rate of 192kHz starts to reach some of the theoretical limitations of modern ADCs. These theoretical limitations do not come into play at lower sample rates.

[Teen Age Riot]

Slightly off-topic question:

How do you determine how good an SRC is? Resample sine waves and compare the artifacts?

[page]

Slightly off-topic question:

How do you determine how good an SRC is? Resample sine waves and compare the artifacts?

I'm lazy and compare existing test results.