Should I record at 44 or 96khz?

[Hugh Robjohns]

Or make a file like I did which doesn't have those artifacts. (snip) It should be pretty simple to take the file I uploaded and plot the actual graph it creates.

Fair enough. Here is a spectral display of your 'misrepresented' file:



This is a screen grab using Izotope RX in spectral display mode. Time is along the horixontal axis, and frequency is the vertical axis, rising to about 96kHz at the top (to show the full audio bandwidth of a 192kHz sampled signal).

Clearly, these are not sine wave tones -- there are spectral components all over the place, almost certainly due to gross aliasing!

The screen resolution and jpeg-ing probably makes it hard to see, but looking at the four tones in your sequence, the lowest spectral components are at 3kHz, 1.5kHz, 6kHz and 1.5kHz again. Assuming the source signals really were sine waves at 18, 19, 20 and 21kHz, these sub-fundamental components would support the suggestion that there is gross aliasing going on.

However, closer inspection of the frequency region between 17 and 22kHz (see below) reveals that none of the four tones in this sequence has any frequency component at 19kHz or 20kHz (although there are frequncy components at 18kHz in all four, 21kHz in the first, second and third, and 19.5kHz in the second and fourth).

This irrefutably indicates that the source files are not as you described at all. An accidental error, or complete incompetence? I'll leave it to the forum members to make up their own minds.

Put your wave files where your mouth is, redcoat.

Okay. Here is a spectral display of a file which I created to do what you said yours was supposed to. Created in Adobe Audition 3, sampled at 192kHz, containing a sequence of true sine waves at 18, 19, 20 and 21kHz, each lasting about 3 seconds and all at a level of -18dBFS (to reduce the risk of frying tweeters).



Again, time is on the horizontal axis, frequency is vertical up to 96kHz, and as you can see, there are just the four spectral lines showing true sine wave tones at 18, 19, 20 and 21kHz.

Should you want to try the wav file yourself, it is here for download (it is 8MB)

If you process this file with any sample rate converter to render a 44.1 or 48kHz file, you'll find all four tones preserved, intact and (hopefully) completely undistorted. I've tried it with SRCs in Wavelab 6, Izotope RX and AA3, all without any problem or significant degradations.

Typically most SRCs will impose a slight (less than 0.5dB) attenuation of the 21kHz tone -- it all depends on the way the filter is designed around the turnover point.

One way or another, you should be able to do something more than sit around misspelling the word color in this conversation.

How witty.

It would appear, from this and numerous of your other posts, that you are far less experienced and knowledgeable than you profess to be and/or may enjoy a degree of 'trolling.' While I (and others here) will be happy to help you redress the former, the latter won't be tolerated.

Hugh

[Shingles]

Assume the Shannon/Nyquist theorem only eliminate aliasing, so your sidebands are eliminated. But it does not guarantee anything about your sample being a great reproduction.

Erm. Yes it does.

[BubbleButt]

this is just a random idea, not based on any scientific analysis, but maybe a better test of 44 vs 96, rather than blind A-B tests, would be a sort of "sustained listening" test ?

ie, is listening to stuff at 44.1 more fatiguing than 96, over a period of hours ? do your ears (or your brain) tire from having to reconstruct those missing frequencies?

maybe those ultrasonic frequencies which we can't actually hear (and which are better captured/reproduced @ 96kHz) have a more long-term or cumulative effect?

a straight A-B comparison doesn't really reflect real-world listening behaviour - for example, i can't imagine most casual listeners quibbling over whether they can hear more detail in a cymbal in two different recordings, however i would expect them to maybe switch a CD off after awhile due to a subconscious "ear fatigue" effect - something that i feel has increased since the advent of CD / MP3 etc vs LP / cassette (although maybe it's just as i've got older i can't handle listening to music the same way ... )

[molecular]

Assume the Shannon/Nyquist theorem only eliminate aliasing, so your sidebands are eliminated. But it does not guarantee anything about your sample being a great reproduction.

Erm. Yes it does.

Have just read the article Hugh linked to earlier on - sorted my head out on a number of issues. One of these was the idea that increasing the sample rate is a step up in 'resolution' akin to moving from DV to HD. I now (feel, at least!) as if I understand how the reconstructed analogue sound that comes out of my speakers is reconstructed identically from either a 44.1 recording OR a 96 recording. (disregarding any 'non-ideal DA' stuff).

I understand the little I do by reference to thinking of the sample rate as storing a 'carrier signal', rather than a 'pixelisation' of the sound. Secondly, from the fact that a dithering process creates a randomly noisy linear signal, in place of a distorted one.

A couple of questions on it, though:

1. The 'dithering process' in question here - in order for the result to be a noisy analogue signal, where the dither noise is what is used to join up the steps in reconstructing the waveform, I am assuming that all DA converters add an *analogue* dither noise after the distorted/stepped signal has been created. Is this right?

2. I had only thought of dither in terms of word length conversion, as a noise that was added to force information regarding low amplitude detail further up the word (?! sorry...) so that it is not just chopped off. Are these two distinct uses of dithering, or am I missing the connection between the two?

Thanks,

Hector.

[Hugh Robjohns]

this is just a random idea, not based on any scientific analysis, but maybe a better test of 44 vs 96, rather than blind A-B tests, would be a sort of "sustained listening" test ?

Certainly listening fatigue is a useful indicator of audio quality -- and it is one area where mp3s and other data reduced formats tend to fall down.

do your ears (or your brain) tire from having to reconstruct those missing frequencies?

In the case of many data reduction systems,the asnwer is yes. However, many decades of using reduced bandwidth systems like 78rpm records, AM radio, cassette and others would seem to suggest that los bandwidth isn't such an issue -- the fatigue problem tends to manifest when small spectral chunks are missig from within the bandwidth (which is what heppens in most data reduction algorithms).

maybe those ultrasonic frequencies which we can't actually hear (and which are better captured/reproduced @ 96kHz) have a more long-term or cumulative effect?

Maybe. Personally, I think perceived differences are more to do with the small phase differences caused by the anti-alias and reconstruction filtering that occur within the 20Hz-20kHz region for base rate and higher rate sampled systems.

Hugh

[Hugh Robjohns]

Have just read the article Hugh linked to earlier on - sorted my head out on a number of issues.

Glad it helped.

1. The 'dithering process' in question here - in order for the result to be a noisy analogue signal, where the dither noise is what is used to join up the steps in reconstructing the waveform, I am assuming that all DA converters add an *analogue* dither noise after the distorted/stepped signal has been created. Is this right?

No, dither is not applied in the D-A process. It is applied when a signal is first quantised in the A-D, or re-quantised to reduce its wordlength. In effect, the dither signal is added (mixed with) to the input audio signal, and the whole thing is then quantised. As Dan explained earlier, the noise forces the signal to jump between adjacent quantising levels in a way that, statistically, represents its true amplitude.

2. Are these two distinct uses of dithering, or am I missing the connection between the two?

They are separate applications of dither, and the technique used differes slightly between the two processes -- but the underlying concept is the same: dithering linearises the quantising transfer function, providing a (theoretically) perfectly linear system with a defined noise floor.

In an A-D converter, the dither signal is am independent noise source of some form. In requantising to reduce word length, the dither signal is usually derived from the lower order bits that would other wise be discarded in the truncation process.

Hugh

[AngryMonkney]

Hi,
I am the original poster of the question, asking if I should use 44 or 96.

I have read this thread and decided to use 24bit/44.1.

I would however like to give a massive thanks to everyone who has contributed, I really appreciate the efforts everyone has gone to.

Thanks,
AM.

[Mike Stranks]

... and relax...

[Interweaved]

Or make a file like I did which doesn't have those artifacts. (snip) It should be pretty simple to take the file I uploaded and plot the actual graph it creates.

Fair enough. Here is a spectral display of your 'misrepresented' file:

[/quote]

Well.. your image didn't load for me. But I'll take your word on it that it's distorted, really. Don't know why it did that (I just downloaded some source code that says it generated a sine wave).

Truth is, I have this example that convinces me using another technique, but I don't think recording it will work nor convince anyone, because it adds another "sampling" layer when recording it.

So.... anyhow, I'm not trolling. As for credentials, I really didn't say much about them.

I'm done with screwing around with the question, because, like I said, not much I'm doing has much energy above 15kHz anyhow.

If you can generate high frequency (17kHz plus) sine waves at 192 kHz/24 bit sampling... then try and play them back at different sampling rates, I think you should hear a difference. Up to you, though...

In fact, I recommend playing up a major scale, first from 8 kHz, then from 16 kHz.. with sine waves... created/sampled at 192/24... then playing it back at different sampling frequencies.

[Hugh Robjohns]

... and relax...

hugh

[molecular]

thanks hugh for these answers.

I realise I'm drifting ever further away from the posters topic, but he seems to have come a wise decision - certainly the same one I have come to, as I am running a laptop and don't want to be the one to say 'no you can't double track that harmony... I'm sampling at 96khz for reasons I can't hear, and I'm maxed out!'

I want to run my understanding about averaging/dithering at AD stage past the forum...

So, I have a sine wave at 100hz, of an amplitude of the first single step of my wordlength (i'll call it '1'...), and I'm sampling at 44.1 - this means that the first quarter cycle, which rises linearly from 0 to 1, is covered by 1100 and a half samples.

What I end up with in the digital domain (once dithered and sampled) is that the first few of these samples are most likely to be zeroes, but as we move through the samples, they become more and more likely to be 1s, following a graph of 'probability of being 1' against 'time' which looks very similiar to my original sine wave...after sample 1101, the likelihood of a sample being registered as a 1 decreases again, until the cycle moves into its negative phase.

Without the dither, I would get ALL zeroes, followed by ALL ones, followed by ALL zeroes in the first half of the cycle.

Is this right? apologies for the gobbledegook explanation, but it is 1am!

Ta,

Hector.

[Hugh Robjohns]

Well.. your image didn't load for me.

Apologies. It was referenced from an in-house file store that may not be visible to ordinary forum users. I have now re-referenced the file to an external web host, so it should be properly visible to everyone. Sorry about that.

If you can generate high frequency (17kHz plus) sine waves at 192 kHz/24 bit sampling... then try and play them back at different sampling rates, I think you should hear a difference. Up to you, though...

I can, as I have demonstrated -- and so too can most people with a half decent DAW -- but if you play a 192kHz file back at a lower sample rate without sample rate conversion -- yes, some DAWs and wave file editors will let you do that -- you will simply lower the frequency of the source tone in a simple 'varispeed' manner.

So what you really need to do is sample rate convert the source file (I'm sure you meant that) -- but then the results will depend entirely on the design and efficacy of the SRC's anti-alias filtering -- and pretty much every SRC does it slightly differently. Inherently, the nearer the source tone is to the notional turnover frequency (which is typically going to be between 20 and 22kHz) the more its amplitude is likely to be reduced, and the more likely it is that some harmonic aliasing might occur (depending here on the harmonic precision of the source audio).

As I explained in my post above, I did actually SRC my own test file described above using the SRCs including within AA3, Wavelab 6 and Izotope RX pro (this last with several different filter options).

In some cases I couldn't hear or measure any differences at all, but in others I could -- but I'm not surprised by that. No one has denied that practical anti-alias and reconstruction filters can, and often do, have a sonic impact on HF signals. You would expect them to because of the inherent amplitude and phase response wobbles as you approach the turnover frequency.

As I explained several pages ago, that is the precise reason that operating a poorly designed converter at a 96kHz sampling rate can often sound subjectively better than when using it at 44.1. It ain't rocket science.

But of course, none of this comes remotely close to justifying your earlier statements and claims relating to futile 2F sample rates and your apparent ignorance or misunderstanding of the Nyquist theorum.

In fact, I recommend playing up a major scale, first from 8 kHz, then from 16 kHz.. with sine waves... created/sampled at 192/24... then playing it back at different sampling frequencies.

Easily done, but time consuming and I'm out of the office all day tomorrow. Why not redeem yourself and impress us all with your technical skills in generating such a potentially interesting test file that we can then all enjoy experimenting with?

Hugh

[Hugh Robjohns]

What I end up with in the digital domain (once dithered and sampled) is that the first few of these samples are most likely to be zeroes, but as we move through the samples, they become more and more likely to be 1s, following a graph of 'probability of being 1' against 'time' which looks very similiar to my original sine wave...after sample 1101, the likelihood of a sample being registered as a 1 decreases again, until the cycle moves into its negative phase.

Yes, essentially. The statistical probability is as you have described, and the degree of 'randomness' between adjacent quantising levels of adjacent samples is the element we hear (and measure) as the noise floor.

Without the dither, I would get ALL zeroes, followed by ALL ones, followed by ALL zeroes in the first half of the cycle.

Yes. Precisely. There is no noise floor -- the system is perfectly silent -- but the wanted signal ceases to be audible until it reaches a level when the quantiser can switch between levels -- and then it is horrendously distorted because it's amplitude envelope is essentially being clipped into square waves.

You can hear this effect very clearly in the 3-bit examples from my february 2008 digital myths article.

I took the analogue output from a CD player reproducing some piano music and passed it through an A-D converter modified to operate with just three bits and no dithering.

truncated - 3 bits

What you hear when the piano is played loudly is gross quantising distortion, rendering the audio very unpleasant and barely recognisable. As the piano tones die away on sustained notes, they break up and then mute, as the quantiser is forced to generate all zeros.

However, when correctly dithered, the piano is distortion free, even during the die away of sustained notes -- but the noise floor is ridiculously high, as you would expect in a three bit system.

dithered - 3 bits

The third file uses noise-shaped dither to move much of the spectral elements of the dither signal to the extreme HF, where is is subjectively less objectionable and intrusive.
Noise shaped dither - 3 bits

The result is a surpsingly listenable version of the piano track, distortion free, with the tail end of sustained notes fading gracefully under the noise floor.

So yes, you are right and I hope these examples cement the ideas for you.

Hugh

[Interweaved]

So what you really need to do is sample rate convert the source file (I'm sure you meant that) -- but then the results will depend entirely on the design and efficacy of the SRC's anti-alias filtering -- and pretty much every SRC does it slightly differently. Inherently, the nearer the source tone is to the notional turnover frequency (which is typically going to be between 20 and 22kHz) the more its amplitude is likely to be reduced, and the more likely it is that some harmonic aliasing might occur (depending here on the harmonic precision of the source audio).

Uh, this I think was the "distortion" I was complaining about... amplitude loss and frequency inaccuracies at high frequencies.

Anyhow, as per doing something more with that, perhaps I'll come back to it in a few weeks. I mess with this stuff in my spare time, I have some other priorities.

[Hugh Robjohns]

Uh, this I think was the "distortion" I was complaining about... amplitude loss and frequency inaccuracies at high frequencies.

A small reduction in amplitude is certainly possible -- but in most cases we are talking about fractions of a dB. Not only can most people not perceive that small a level difference, but most people won't be able to hear a 21kHz sine tone anyway.

I'm not sure what you mean by 'frequency inaccuracies' The pitch of the source tone can't change through sample rate conversion, and if the source tones were pure sinewaves there will be no harmonics to cauase aliasing distortions. So if you can hear additional frequency components, that would suggest an imperfect source signal combined with inadequate anti-alias filtering. Both are possible.

So if they are hearing some significant and repeatable difference, the liklihood is that they are hearing some kind of non-linear distortion products or aliasing -- which inherently means that the source signals weren't processed properly with an accurate anti-alias filter before digital conversion (or sample rate conversion), and/or the replay system is inadequately engineered to reproduce HF signals without adding significant distortion products of its own.

Neither of these things is inherent in digital recording/processing per se, but are possible through poor equipment designs. An important distinction to make.

Anyhow, as per doing something more with that, perhaps I'll come back to it in a few weeks. I mess with this stuff in my spare time, I have some other priorities.

As do we all. I hope your understanding of this complex topic has improved through our dialogue

Hugh

[Richard Graham]

Hugh displaying (as usual) the patience of a saint.

Is it too late to join the debate and answer the original question, i.e. 'Should I record at 44 or 96kHz'?

I think the short answer is 'YES'.

As in, yes, you *should* record at 44.1 or 96kHz.

Admit it, Angry Monkey, it *was* a trick question, wasn't it?

[ken long]

Fantastic read and kudos to the admin for dedicated insights and thorough explanations.

Just wanted to add:

If you are recording wildlife, it is advisable to record at the highest possible rate. Just because human perception is limited to the 20Hz-20kHz range (debatable itself), doesn't mean there isn't information above that.

ken

[muzines]

Not only can most people not perceive that small a level difference, but most people won't be able to hear a 21kHz sine tone anyway.

My ears top out at about 17.5KHz...

[molecular]

If you are recording wildlife

you mean if you are recording FOR wildlife!

[Richard Graham]

Fantastic read and kudos to the admin for dedicated insights and thorough explanations.

Just wanted to add:

If you are recording wildlife, it is advisable to record at the highest possible rate. Just because human perception is limited to the 20Hz-20kHz range (debatable itself), doesn't mean there isn't information above that.

ken

Especially if you are recording *for* wildlife. Bats in particular are picky about sampling rates, as anything lower than 96k causes them to crash into a tree.

[Richard Graham]

Molecular got there first! Bah!

[ken long]

If you are recording wildlife

you mean if you are recording FOR wildlife!

I don't get it.

[The Korff]

I think he means, if the intended audience of the recording IS wildlife, then the little animals will perceive the difference between 44.1kHz and 96kHz recording much more clearly (though this also depends on other factors, such as the extended frequency response of the microphones used, and the speakers that you use to play your recordings back to the animals with).

I've ruined it, haven't I...

[muzines]

I don't get it.

Bats are really unappreciative of low bit-rate mp3 compression artifacts...

[molecular]

I'm no expert on this, but the two types of bat common up here are identified with a detector that pitch shifts their echo-location signals. As far as I can remember, the species are identified by audio at around 40khz and 60khz respectively.

I don't know if digital versions of these are widespread... I'm guessing there's no reason for it.

I hope this answers the OP's question!!!