Benefit to go with 96k?

  • Thread starter Thread starter frist44
  • Start date Start date
Teacher said:
u missed what blue bear was saying (i think). He's saying there's no such thing as recording at 32 bit because no interface yet supports it.
Click to expand...
I would argue that no interface will support it ever! (Not in our meager lifetimes anyway.) 24-bit is just hard enough as it is. If the entire dynamic range and linearity was utilized the precision would be better than that of the surface of the mirror in the Hubble (or any other current) telescope! There is a dynamic range of 144 dB (give or take a few millibels) in a 24-bit sampling!

On the other hand I would argue that 32 bits is not nearly enough when you start doing calculations (effects)...
 
Thanks a lot Slack and Bear. I really appreciate your taking time to help me out. I will resume recording as normal in the safe knowledge that I know a little more about my system now ;)
 
BasPer said:

I would argue that no interface will support it ever! (Not in our meager lifetimes anyway.) 24-bit is just hard enough as it is.
Click to expand...

Look at how fast computer technology has developed over the last twenty years. In 1982 we could only dream of the possibilities that we have now. Could we imagine it at all that we would be able to mix dozens of tracks on something as small as a laptop? That's the reality of today.
Hell, did we know what a cd was back then? The whole world was surprised that all that music would fit on such a small disc. And you didn't even had to turn it to hear the other half of the album! And now we have DVD's...

I'd say that we'll have 32 bits interfaces. Well before the end of the decade.
 
christiaan said:

I'd say that we'll have 32 bits interfaces. Well before the end of the decade.
Click to expand...
You wanna bet? :) Well, OK, we might have 32-bit interfaces some day, but it will be purely a sales gimmick and if they sound better than a 24-bit interface it will be because of other things than the bit depth. I see you are an engineering student, so I callange you to build so much as a sample-and-hold circuit or a buffer stage (a x1 amplifier) with 144dB of signal/(THD+N)! :D
Computers have indeed gotten better, and thats because digital CMOS is childs play, but analog electronics sure is, and has, not! :) (I should probably have rephrased that sentence...)
 
christiaan: 24-bit resolution gives a theoretical noise floor around the level of atomic noise. For the application of A/D or D/A converters, there's no need to go any farther because it wouldn't do any good. Even with really good 24-bit A/D converters, you're only getting probably 22 bits of resolution. Internal thermal noise is all that's populating those bottom couple of bits.
 
pglewis said:
... See, Nyquist says that half the sampling frequency is the highest frequency that can be unambiguously represented. That doesn't mean it's represented perfectly by any stretch of the imagination. In fact, anything at exactly half the sampling frequency (22,050 for 44.1) can only be represented as a square wave since you're only going to get a resolution of 2 samples per cycle....
Click to expand...

Not quite. It would be represented as a sine wave, not a square wave.*

The key element to the whole analysis is that every wave can be represented as a combination of sine waves of various frequencies, amplitudes and phases. Any wave can be unambigously -- perfectly accurately -- represented, so long as either (i) the wave doesn't have a component with a frequency of half the sample frequency or higher or (ii) you ignore whatever components of the wave have a frequency of half the sample frequency or higher. From what I understand, experiments indicate that our ears do the latter at something in the range of 12k to 20k, depending on the ear: that is, ignore the components that have a frequency higher than that. It may or may not be worth pointing out that speakers do the same thing, at some point (though, depending on the speaker, it may be above 20k).

A square wave has a lot of high frequency content -- a square wave at frequency "x" is a sine wave at frequency x plus a long series of sine waves at the harmonic freqencies of 3x, 5x, 7x, etc. If you take a 20k square wave, and cut out all the content that's over 20k, you'll get a 20k sine wave. What's more interesting is that you won't be able to tell the difference, unless you hear frequencies well over 20k.

... At 44.1, we're getting a resolution of less than 4 samples per cycle even down at 12kHz. The big advantage I see to 96k or even 192k, conceptually, is better accuracy with the frequecies we can hear....
Click to expand...

The only way the accuracy at 96k is different is that it (accurately) includes more of the high-frequency content of the signal. But that's irrelevant accuracy, because our ears can't perceive the difference between the "inaccurate" and the "accurate" signal. At least if you accept that we don't hear frequencies above, say, 20k (which means, in this context, not only that we don't hear pure sine waves above 20k, but we don't hear the above-20k component of complex waves with a lower fundamental frequency).


*Also -- as I think you realize -- you can't unambiguously represent a wave at exactly half the sampling frequency, because you don't know where in the cycle your samples are "hitting" -- you could be sampling the peak each time, or halfway up to the peak, or anywhere else. You can't tell the amplitude. You need to be less than half.
 
so still, all this theoretical audio engineering talk, and no one has answers to my questions?
 
yes, higher frequency recording will leave slightly superior representations of some of the higher, potentially audible frequencies. no, you may not hear it-or you may. if you do so, it would probably be best to record at 88 rather than 96, for the much more mathematically simple downsample-think of the simplicity of 1:2 versus the complexity of 1:2.176870748. now, that result is only to 9 digits, and no repetition. imagine what a crazy number like that looks like when represented in blunt digital data while trying to properly downsample. i'm sure you can see what i'm getting at. my recommendation-do you have the equipment? try it at 88, and see if it seems to sound better. better equipment will be much more likely to represent this difference, much better equipment, the kind that if you could afford it, you wouldn't think twice about using 88. your milage may vary, but there are tons of people who swear by the differences. MAYBE, just maybe, if you are doing some digital changes (eq, normalize, effects, etc), this will make a noticable quality difference. maybe it won't. if i had the option, and it didn't kill me with performanc slowdowns, i would do it.

in relation to other posts about 32-bit, my understanding is that with electronics as we know them, it's impossible, even in theory, to get more than about 29 bits of non-noise data-by non-noise, i mean not including noise created by the CLEANEST link in the chain-the actual dsp chip. even if that were the case, you probably wouldn't get better than about 24 or 25 once it goes through anything else, but at LEAST it might be better than the max of around 22 on 24. that said, folks like neve are working on more bits, but it's a slow-going process. i think that if a totally new transduction system can be used (optical circuitry, for example), it may be possible to approach that magic 32, but even then, who will be able to hear it? noone, most likely. 24 bits is pretty kickass, and those of you who are even getting a legitimate 18 are really going to benefit from it.
 
i personally think that the extra bits should be used in a manner that beefs up the resolution of the existing signal, not just increasing dynamic range. what i mean by that is that a 32-bit or higher system should use its extra bits to allow for more variations within the 120db or so that 24 bit gives us, so as to give a better representation of the signal. as in, basically a representation of amplitude to more decimal places than is currently possible. or is that idea useless/already implimented/improbable etc?

same thoughts on higher sample rates, though they already provide this benfit: decreased aliasing and "smoother" waveforms at higher sample rates are, as i see it, the biggest advantage. if nyquist at 44.1 is already 22Khz, I don't think the benefit of higher sample rates come from nyquist at all, with the exception of easier-to-design filters.

anyways, another question:

I have a MOTU 1224. It can record at a max of 24/48. I'm trying to decide whether to track a project at 24/44.1 or 24/48. The project will eventually end up on audio CD, so I know that recording at 44.1Khz to begin with saves from having to do the evil 48->44.1 conversion. However, when I do a mixdown, I'll likely be mixing down to 32/88 or 32/96 (which is better?) to apply final EQ, limiting and dithering, then sample convert down to 44.1

So, my question... what's the optimal sampling rate to record at, and the optimal sampling rate to mix down to? Right now I'm thinking it might be best to record at 48 for best inital quality, then mixdown to 88 for a better later conversion to 44.1

After all, you only lose quality doing a downsample, right? So, it seems logical to me that, because it's first being upsampled to 88 from 48, then simply divided by 2 (as you said) for the conversion to 44.1, that should give the best overall quality.

Thoughts?
 
I don't remember if it has been said further above, but the real benefit in 96k vs. 44.1k is the much simpler anti-aliasing filter that can be used. The brick wall kind of filters often seen in 44.1 AD's because they want to boast 21kHz of bandwidth are by no means inaudiable. In a 96k (or 88.2k because I agree with zer0 about the downsampling) a much less drastic filter can be used, with extremely little effect in the audiable range.
 
bleyrad said:
After all, you only lose quality doing a downsample, right?
Click to expand...
No, upsampling from 48 to 88.2 will decrease the quality. But from 44.1 to 88.2 will not. I suggest that if you are going to do everything yourself then track, mix and master at 44.1, but if you are going to send it to mastering then track and mix at 48 and then let the mastering house convert to 44.1!
 
i'm assuming you said that because there's some really good sample convertors out there that mastering houses most likely have, and with which it's probably better to do the 48->44 conversion than to track at 44 to start with.
is there any software/plugin convertors you know of that do a really good job of this, to the point where it's worth it to track at 48?
 
...you could be sampling the peak each time, or halfway up to the peak, or anywhere else. You can't tell the amplitude. You need to be less than half.
Click to expand...
Actually, this is a good point that I didn't take into account (I know enough to shoot myself in the foot, but my experience with ones and zeroes is as a developer; not an audio engineer). But it makes perfect sense, now that you've pointed it out.

It still seems to me that there is an appreciable amount of quantization distortion of high-frequency content that is still well in the audible range at 44.1kHz. Even 8k frequency components get sampled with less than 6 points per cycle at 44.1.

Now, I've been very happy with 24/44.1 for what I'm doing. I'm getting better quality recordings than ever before. But I'm sure there are plenty of people who can consistently pick 96k recordings out from 44.1k recordings in a double-blind. Is this because of content > 20k? Perhaps, but I still seem to think that the cumulative quantization distortion of frequency content below 20k has more to do with it than the dog-whistle range.

This is just because it makes sense to me not out of any practical application, so this is an interesting discussion to me. Especially since there are a few guys around, like you, who really understand the theory behind it.
 
bleyrad said:
i personally think that the extra bits should be used in a manner that beefs up the resolution of the existing signal, not just increasing dynamic range.
Click to expand...

It's the same thing.

Here's a perhaps-intuitive way to look at it. When you try to represent a figure with a limited number of bits, each measurement will be slightly off, because you've got to round measurements to the nearest discrete value that you can represent. These errors are "noise." The smaller the errors are, the higher the ratio of signal to noise, and the higher the ratio between the largest signal you can clearly represent and the smallest signal you can clearly represent. Beefing up resolution increases dynamic range.

On the question that started this whole thing, I tend to think there's something to what BasPer says. There is probably some advantage to moving your anti-aliasing filter way up out of the audio range. Whether it's worth the extra data depends on a bunch of other factors.
 
sjjohnston said:
quote bleyrad:"i personally think that the extra bits should be used in a manner that beefs up the resolution of the existing signal, not just increasing dynamic range."

It's the same thing.
Click to expand...
Yes, and a lot of otherwise knowledgeable people seem not to have fully grasped this. I think I'll write a little something about it, but I'll do it in another thread because it's somewhat off topic and possibly long-winded. Hmm, I might even try to make some educational figures to go with it... (Because I know you guys just love educational Swedish pictures. :D) Well, that wont be until next week sometime anyway since I'll be spending the entire weekend singing Beethoven.
 
pglewis said:
It still seems to me that there is an appreciable amount of quantization distortion of high-frequency content that is still well in the audible range at 44.1kHz. Even 8k frequency components get sampled with less than 6 points per cycle at 44.1.
Click to expand...

I don't think there is any quantization error. An 8k frequency component means a sine wave of 8k. It only takes two sampling points to represent a sine wave (ignoring the problem of aligning the sampling freqency with the sine wave properly... which is why the Nyquist limit is usually stated as below half the sampling frequency, not equal to or below).

So, 6 points per cycle on an 8k signal is all you need to perfectly define the sine wave. Now, perhaps you're thinking is that it's not a perfect sine wave, but has some irregularities which limited sampling points will not accurately capture. ;) And what causes those irregularities? Higher frequency components! :) So, in a sense, you are correct in thinking that a real instrument playing at 8k might not be accurately represented by 6 sampling points, but that is only because there are harmonics created by the instrument that are well above 8k, which add color etc. up to a point (when they become inaudible... or imperceptible, take your pick on that argument). A true 8k signal is perfectly captured by 6 points (or three, or perhaps two).

So, while it is true that using 44.1 kHz sampling frequency means you can only capture sine waves up to 22 kHz, it probably doesn't matter much. If it weren't a true sine wave, then it would be because of higher (than 22 kHz) frequency components, which are probably inaudible or imperceptible anyway.

As pointed out above, your ear just doesn't know the difference between a 20 kHz sine wave, square wave, dirty sine, or anything in between.


And I agree with the statements up above that the antialiasing filter is probably the biggest single benefit of using a higher sampling frequency. The high frequency components captured by the higher sampling rate are probably not audible (or perceptible) for most people on the vast majority of reproduction equipment, but the antialasing can be. I may be wrong, but I think some A/D converters actually sample at a higher rate, apply the filter, and then downsample. Even though there are errors introduced by the downsampling, in some cases they are less perceptible those caused by high frequency filtering.

At least, that's what I seem to recall from my very limited knowledge of digital audio theory.
 
sjjohnston, BasPer, bigus: You guys are slowly pounding some important concepts into my head. I'm gonna re-read some stuff on Fourier transforms and hope it doesn't make my head explode.

Also, I meant to mention that more gradual anti-alias filtering as a benefit to higher sampling rates makes intuitive sense to me as well.
 
96k is better. listen to a piano or mandolin sometime at 48k/96k. you are best off using your computer for tracking only. plug-ins suck. never hit "resample." dithering is lame. track at your highest resolution. run it into a decent analog mixer and do your mix there. that way you can output each bus at maximum resolution. christ you can buy neoteks for under $10,000 now. they are good low price mixers. (don't get the series I though) when you do the final mix record it at release resolution, i.e. do one pass at 16/44.1 and another at 24/96k for a possible future dvd release, do not record it to 24/96 then dither it to 16/44.1. every caluculation degrades your audio. of course you should track and mix to analog anyway. 24/96k still does not approach the quality of a good analog recording, why truncate it, round it off and resample it?
 
You must log in or register to reply here.
Back
Top