Great Article - Explaining why 24bit/192kHz has no benefit.

[theblue1]

I have a quick question from the half-clued in quarter (that would be me)...

Is the reason that Nyquist-Shannon stipulates an infinite process because -- pardon my clumsiness here -- if you start and stop sampling at random points in a signal stream the first and last samples will almost certainly be inaccurately derived because they will almost certainly not experience full wave cycle during the initial and final sample cycles?

Or am I stumbling around in my own nescience again?

[joeq]

Originally Posted by illacov

But by conducting ABX on a phenomenon, you've already got to have some kind of plausibility that the phenomenon is repeatable and therefore quantifiable correct?

my hierarchy would go like this

a) Anecdotal reports of people saying such and such is "better" or even just "different". Rumors.

b) Blind A/B/X testing to confirm that these perceived differences are not just occurring to people who are peeking at the screen. That they are real differences, IOW.

c) A discussion of digital theory and what might be going on down inside the bowels of the DAW that could explain the results obtained in 'b'.

I for one, see no point in having the discussion in 'c', to "explain" the anecdotes in 'a' without going through 'b' first. Include the plug-ins if you like, but don't skip the test.

Please show me where the logic I'm using is so backwards. Would we rather argue about the egg or get to the part where we cook it?

I have no problem with including plug-ins or any other aspect of 'working at' the higher sample rates in the A/B/X testing.

I just think it only logical that b) should come before c), that's all.

I fail to see how this makes A/B/X a 'red herring'.

To skip the testing phase because you think the sound differences are "obvious" is a mistake. If these differences are truly "night and day", the test should take a few minutes administer and be more than conclusive.

If, on the other hand, it is something that people can only 'hear' when looking at the labels, well, as you can imagine I would think discussing theory is a waste of time. Many a night-and-day difference vanishes into nothingness when the blindfold goes on. Just sayin'.

ARE there double-blind listening tests using plug-ins that prove people are reliably hearing the difference?

Links? If so, I withdraw my objection to the theoretical discussion.

[UnderTow]

Originally Posted by theblue1

I have a quick question from the half-clued in quarter (that would be me)...

Is the reason that Nyquist-Shannon stipulates an infinite process because -- pardon my clumsiness here -- if you start and stop sampling at random points in a signal stream the first and last samples will almost certainly be inaccurately derived because they will almost certainly not experience full wave cycle during the initial and final sample cycles?

Or am I stumbling around in my own nescience again?

It is not just the first and the last samples. Look at the shape of the sinc function (for instance in the Lavry paper). It gets closer and closer to zero as it gets further away from it's center point.

Well I need sleep and typing on a phone sucks so unless the hint in the previous paragraph is enough, I'll elaborate tomorrow.

Alistair

[audslu]

Originally Posted by UnderTow

The timing resolution of digital audio can be calculated with the following formula: 1 / ( 2pi * FS * bit depth). For CD audio that gives 1 / (2pi * 44100 * 65536) = 5,5 * 10 ^-11. That is 55 pico seconds. In other words, the timing accuracy of 44.1/16 audio is in the pico second range while human hearing time resolution is in the microsecond range.

CD gives a hundred thousand times more timing resolution than the human ear needs. Also, as you can see by the formula above, bit depth tells you more about the timing resolution than the sampling rate does.

Alistair

So actually you say it is better but makes no difference to the human ear. If it is correct it proves it in theory but no abx test can show.

[illacov]

Originally Posted by joeq

my hierarchy would go like this

a) Anecdotal reports of people saying such and such is "better" or even just "different". Rumors.

b) Blind A/B/X testing to confirm that these perceived differences are not just occurring to people who are peeking at the screen. That they are real differences, IOW.

c) A discussion of digital theory and what might be going on down inside the bowels of the DAW that could explain the results obtained in 'b'.

I for one, see no point in having the discussion in 'c', to "explain" the anecdotes in 'a' without going through 'b' first. Include the plug-ins if you like, but don't skip the test.




I have no problem with including plug-ins or any other aspect of 'working at' the higher sample rates in the A/B/X testing.

I just think it only logical that b) should come before c), that's all.

I fail to see how this makes A/B/X a 'red herring'.

To skip the testing phase because you think the sound differences are "obvious" is a mistake. If these differences are truly "night and day", the test should take a few minutes administer and be more than conclusive.

If, on the other hand, it is something that people can only 'hear' when looking at the labels, well, as you can imagine I would think discussing theory is a waste of time. Many a night-and-day difference vanishes into nothingness when the blindfold goes on. Just sayin'.

ARE there double-blind listening tests using plug-ins that prove people are reliably hearing the difference?

Links? If so, I withdraw my objection to the theoretical discussion.

Well help me to verbalize this.

I think the differences emerge ONLY when you put the higher sample rate audio through plugins. So 44.1khz vs 96khz thru the same EQ (if it doesn't use oversampling) should sound different.

And I think those differences disappear when you use plugins that use oversampling.

So what I'm getting at is that without the necessary devices in place (plugins that don't use oversampling coding) the listeners are not able to hear those differences and there are possibly NO differences.

There ARE actually files on Gearslutz ov Waves SSL at 44.1 vs 96k where there were differences. It was a good while ago but very revealing about my growing feelings/findings about oversampling theory in DSP.

So back to the original issue I'm having, should we ask a group of scientists if they can see cells without a microscope? I think you'd agree that would be pretty absurd. You can't see cells without the proper optics. The same applies to hearing the differences in how different sample rates react within the DAW environment (to create the premise that higher sample rates are better).

In order to observe the phenomenon or measure it, it has to be repeatable. So in this case you have to create the environment for it to be detectable.
You have to create the premise. Without the premise there is no phenomenon is all I'm saying.

If you want my thoughts, I think you can't hear the differences without the premise I've been harping on, so the ABX part is only going to prove what to me is very obvious and already a mute point. I mean if that's the only way people will get around to actually testing within the DAW environment, the way most engineers listen to their raw recordings, then so be it, but its kind of weird to get mired down in something that is to me wrote and only a staple on the bulletin board that holds the term paper that gets the teacher's grade.

It detracts from the heart and more important thing that's at stake here and leaves alot of dogs chasing their own tails and barking at themselves in the mirror.

I mean hey on some level, oversampling or higher sampling rates both have their costs. Oversampling eats up CPU cycles and high sample rates eat up storage space. I'd rather burn up CPU vs storage any day of the week to be honest.

Peace
Illumination

[theblue1]

Originally Posted by UnderTow

It is not just the first and the last samples. Look at the shape of the sinc function (for instance in the Lavry paper). It gets closer and closer to zero as it gets further away from it's center point.

Well I need sleep and typing on a phone sucks so unless the hint in the previous paragraph is enough, I'll elaborate tomorrow.

Alistair

Not to worry. I'll go do some heavy lifting with Mr Lavry's paper. I was being lazy.

[joeq]

Originally Posted by illacov

I think the differences emerge ONLY when you put the higher sample rate audio through plugins. So 44.1khz vs 96khz thru the same EQ (if it doesn't use oversampling) should sound different.

And I think those differences disappear when you use plugins that use oversampling.

I got that.

And so, I ask again why not include these in the ABX testing? Why blame blindfold listening ITSELF, when it is not the testing, but the methodology of the test that is objectionable to you?

You are coming off as being AGAINST the idea of blindfold testing altogether! If you are correct, then a simple blindfold test should be more than enough to be conclusive.

So what I'm getting at is that without the necessary devices in place (plugins that don't use oversampling coding) the listeners are not able to hear those differences and there are possibly NO differences.

Again, IMO this is not valid as an objection to A/B/X testing per se, but it IS valid as a demand to have (non oversampling) plug-ins included in the A/B/X test!

I totally understand your demand to have non-oversampling plug-ins included. I don't understand how you manage to separate the concept of "A/B/X testing" as if somehow it excluded trying a version with these plugs?

[Jon Hodgson]

Originally Posted by UnderTow

Hmmm... First up, how do you arrive at 360nS for DSD? I get 1 / (2pi * 2822400 * 1 ) = 5,64 * 10 ^ -8.

As for the rest, I'm going to have to think about it. :-)

Alistair

Guilty on the first bit, I missed the 2pi term in my quick maths, but the point was that DSD would by definition have lower timing resolution if that formula was correct... in fact I wonder if you can really talk about full scale with a 1 bit quantizer, there is no such thing as a quantization step size in that scenario, the level is either above or below the quantization point.

Anyway, the fact remains that the formula you have appears to be some attempt to work out a possible deviation of a single sample, and from what I can see doesn't really do that properly.

And it certainly doesn't give us the accuracy of the signal capture, the key point though is that the sample isn't important, the signal is. Of course the sample is part of the signal, but you really can't learn that much focusing on single samples.

[Jon Hodgson]

Originally Posted by audslu

Again from wiki Nyquist-Shannon theorem about Shannon's proof:
"As in the other proof, the existence of the Fourier transform of the original signal is assumed, so the proof does not say whether the sampling theorem extends to bandlimited stationary random processes."

So is this true or false? It seems the foundation of digital audio is not so clearly explained... (to me at least!).

Shannon, and we, are concerned with time limited functions, a message (his field was communications) has a start and an end, as does a song.

Basically they are by definition NOT statiionary random processes.

Also Shannon uses FT while in reality we use FFT as an infinite Fourier series is impossible. I don't think it's exactly the same.

No, we use a DFT (Discrete Fourier Transform, of which an FFT is an efficient implementation) because we're dealing with a discrete time signal.. i.e. one that has been sampled.
Shannon use a Fourier Transform because that is the applicable maths in the continuous time domain.

The second signal you say, can it be added itb in any way?

The second signal is the error, you don't get to choose whether or not you have it (though you can manipulate its content through the design of the quantizer).

Is there a common preference that noise is better? I hear everyone saying how good and musical and "warm" is gear with HD all the time. I d say it's a subjective preference despite both being measurable.

Some people like a bit of quantization distortion on certain instrument sounds as an effect, but there's nothing "warm" about it in my opinion, it's pretty well enharmonic in nature, more grit and dirt.

[Jon Hodgson]

Originally Posted by theblue1

I have a quick question from the half-clued in quarter (that would be me)...

Is the reason that Nyquist-Shannon stipulates an infinite process because -- pardon my clumsiness here -- if you start and stop sampling at random points in a signal stream the first and last samples will almost certainly be inaccurately derived because they will almost certainly not experience full wave cycle during the initial and final sample cycles?

Or am I stumbling around in my own nescience again?

The reason you theoretically need an infinite length signal is that you also theoretically need a signal that has zero content outside of the desired band.

Mathematically you can't have this, if it's a finite length then it's always got infinite bandwidth, but in practice this isn't an issue.

[doug hazelrigg]

I'm wondering if anyone anywhere at any time, including audio engineers and audiophiles, ever heard a song and thought, "That song would have been MUCH better if they had only recorded it using 96kHz fs before downsampling it to 44.1kHz

This is total pedantry, gents

[joeq]

Originally Posted by doug hazelrigg

I'm wondering if anyone anywhere at any time, including audio engineers and audiophiles, ever heard a song and thought, "That song would have been MUCH better if they had only recorded it using 96kHz fs before downsampling it to 44.1kHz

This is total pedantry, gents

and yet through it all, I find a pedantic attitude infinitely less offensive than a patronizing one.

[Jon Hodgson]

Originally Posted by audslu

That's what i'm saying, in the proof to the theory that just above the Nyquist rate ALL data is preserved FT is used for infinite time continuous signal, when you quicker transitions (process), you need bigger bandwidth. Ideally infinite bandwidth can constract a perfect square. There is no perfect squares of course in reality, but the higher the bandwidth the closer you can get to sample and reconstruct that perfect square with less pre-ringing and ringing?

We don't need to capture a perfect square wave, we need to capture those parts of a square wave that we can hear (generally thought to be those below 20kHz).

But the error is the signal that you tweak itb right? It doesn't go straight to dac to reconstruct with the same algorithm (src, filters etc).

The error is part of the signal, it gets processed along with it, exactly like when you put the signal through an analogue stage and some noise is added, that noise is then part of the signal and gets processed along with it.

So if i understand correct the rest FFT which happens itb is based on that signal.

What's with the constant reference to Fourier Transforms? They don't figure in this, they're a tool used for analysis and manipulation, not part of the sampling process.
If we want to convert from the time domain into the frequency domain, or vice versa, we use a fourier transform. That's useful if we want to analyze what is in the signal, or to manipulate it in various ways, but that's totally seperate from sampling and playback.

The set of samples is the signal, it's not something that represents the signal if you give it a twist and a shuffle, it is the signal.

Perhaps it's easier for you to grasp this if you think about a bucket brigade delay line (an analogue delay, you've probably at least encountered a guitar pedal using one at some point).

It is a sampling system, only there's no quantizer, the levels are captured by a capacitor, and then passed along the line of capacitors with each sample period (like a bucket brigade passing buckets) to the output.

If you just hook your amp up to that output, you'll hear your orginal signal. Now, if you didn't bandlimit the input you're going to hear your original signal plus aliasing noise because the process of sampling introduces aliasing noise if the signal isn't bandlimited and if you Nyquist frequency is below your audible limit then you're going to hear stuff above your original signal because a discrete time signal has repeating spectrl content (that's why we need an anti imaging filter on the output), but your original signal never went anywhere, and it's stil there.

The more you process itb, the more you drift away from what the analog non quantised signal would be, with similar analog process eg compression.

Not really

nyway just my thoughts, i don't understand digital audio that much but i think the smaller part of a Fourier series we use to manipulate audio in realtime the further away we go from the purely theoritical infinite Fourier series which proves the Nyquist-Shannon theory.

I'm afraid it's quite clear you don't understand samplng, the problem is you're trying to then make reasoned judgements and extrapolations about somethnig you really don't have any understanding of.

Perfect conversion to me is infinite bandwidth

And straight away you're into what we're not trying to do, we're trying to capture what we can hear

zero jitter (oversampling falls into jitter),

Oversampling is not jitter, you really need to do a lot more learning and rather less thinking at this point

infinite slew rate in the analog part, which doesn't exist.

And is only necessary for infinite bandwidth

So no data loss is an audiophile myth. Good converters are good enough for me, but itb there is big number of proceses going on, not adc and straight to dac.

Depends on what you're doing with the signal.
And nobody has ever claimed no data loss (or more precisely information loss).
The question is not whether we can capture everything perfectly, it is whether we can capture what we hear as near to prefectly as we cannot hear the difference.

[doug hazelrigg]

Originally Posted by joeq

and yet through it all, I find a pedantic attitude infinitely less offensive than a patronizing one.

A fair response.

But... infinitely?

[Jon Hodgson]

Originally Posted by joeq

and yet through it all, I find a pedantic attitude infinitely less offensive than a patronizing one.

So, does this mean that being pedantically patronizing balances out?

[kennybro]

Originally Posted by doug hazelrigg

I'm wondering if anyone anywhere at any time, including audio engineers and audiophiles, ever heard a song and thought, "That song would have been MUCH better if they had only recorded it using 96kHz fs before downsampling it to 44.1kHz

This is total pedantry, gents

My guess would be, yes, it's been thought and said often. Which brings us back full circle to the first page.

[joeq]

Originally Posted by doug hazelrigg

A fair response.

But... infinitely?

OK maybe not infinitely... but in all seriousness, should not EVERY song benefit, if only ever so slightly, from being recorded "better"?

Many of the Technical Types are engineers recording the Song The Band Brings In. If such people want to delve into the finer points of the technology, it doesn't make that song worse.

Unless the songwriter himself is taking time off from his music lessons to discuss sample rates... But then, said songwriter should not even be pursuing audio engineering at all, lest it distract from 100% focus on the "good song"!

[theblue1]

Originally Posted by doug hazelrigg

A fair response.

But... infinitely?

I keep lecturing you guys on properly qualifying your statements...

[doug hazelrigg]

Originally Posted by joeq

OK maybe not infinitely... but in all seriousness, should not EVERY song benefit, if only ever so slightly, from being recorded "better"?

I'm something of a hypocrite -- check that -- I'm a TOTAL hypocrite -- because I've spent countless hours debating this subject both here and elsewhere. In general, I'm convinced using very high sample rates is not only pointless, for numerous reasons, but possibly even detrimental -- these are physical machines, after all, and they have real-world, insurmountable limits that have NOTHING to do with the current state of technology, etc. (although there are some medical applications that oversample at rates that will blow your mind)

But none of that is my CHIEF frustration on this issue... that would be that, in my mind, there is a glaring disconnect between us pro and semi-pro audio guys when we assert that higher sample rates yield more "accuracy" and therefore higher "quality"... and in the next breath claim that "analog is better." Those are NOT consistent assertions!

In any case... having examined this subject for some time now, and having discussed it directly with guys like Lavry, Aldrich, and Massenburg, I am of the position that:

>although sample rate is related primarily to frequency, it DOES have a relationship with both phase and amplitude

>if you go through the math/trig involved, it can be shown that various sample rates yield differing errors in computing peak values; as one increases sample rate, these errors diminish

>there's a LOT more to digital audio than understanding Nyquist-Shannon, and a lot of people who cite it do so incorrectly...

>it's a FACT of nature: almost all high frequency (>2B) in the type of signal we will typically sample is of a relatively low energy... claiming we need to capture those frequencies in order to accurately replicate the original signal (because we supposedly somehow "perceive" them) is quite dubious

[Syncamorea]

Originally Posted by Jon Hodgson

So, does this mean that being pedantically patronizing balances out?

Only if the pedantically patronizing people are proportionally proper in their perspectives.

[Jon Hodgson]

Originally Posted by audslu

What's with Fourier?? Obviously you don't understand digital audio. What do you think plugs and algorithms do, process time? How softsynths generate waveforms, they put samples in your screen to watch? You don't have a clue my friend and i was wrong to ask you it seems.

Erm, you mean plugins like this one?

Instrument Overview | GFORCE SOFTWARE

Would you like to take a guess who programmed it?

All of it, most especially the audio code (no libraries there)... well bar a couple of handy little snippets supplied by Laurent De Soras.

And also analyzed the original hardware OSCar schematics, probed the actual synth and listened to the output (and various tapped points in the audio path) to develop the algorithms that were then coded.

[squirreltrench]

Originally Posted by Syncamorea

Only if the pedantically patronizing people are proportionally proper in their perspectives.

I did an A/B/X test on pedantically patronizing people, and they did indeed null out.

[Jon Hodgson]

Originally Posted by audslu

Ok good to know.

That's all you have to say after saying I have no clue about digital audio?

I have considerably more than a clue, as the results of my work in digital audio confirm.

You on the other hand, have a number of notable misconceptions, as your last but one post confirms.

I like oddity and symptohm very much. Is there any chance GF will make v2 of oddity with a better non-stepping filter?

Oddity and Symptohm are both products the Ohnforce guys, whilst I've had many an email conversation with several of them about programming issues, I have no knowledge of any of their project plans, so I couldn't say.

[Ethan Winer]

Originally Posted by Jon Hodgson

Instrument Overview | GFORCE SOFTWARE

Nice one Jon!

One of these days I'd like to pick your brain about some of this stuff. Are you in the US?

--Ethan

[audslu]

Ok chill. With all due respect Mr Hodgson, but it seems we don't understand each other very well and i'm terribly sorry about my lack of the english languange terminology, but lets don't play with semantics. I don't claim to be an expert, i have already stated that. I like conversations about digital audio, and i might also learn something new from this. With you saying"Oversampling is not jitter" i thought "no kidding?" ... what is this guy talking about?

About oddity filter if you move the cutoff fast enough will you hear what i'm saying. Another example is pro53 filter, more noticeable there. Newer softsynths don't do that, and i'm not talking about diva only.

I was kind enough not to mention i don't like imposcar (1 or 2), never liked it, sorry. Pure coincidence or not, that's another topic.

I have heard many softsynths (including freeware) as good or even better made with synth-edit, it doesn't mean these guys are digital audio experts neither do they build converters or design converter chips! Programming a synth does not prove digital audio expertise to me nowadays. Also there is freeware parts of code and opensource projects out there, so one can copy and learn code along the way.

[theblue1]

Originally Posted by Jon Hodgson

Erm, you mean plugins like this one?

Instrument Overview | GFORCE SOFTWARE

Would you like to take a guess who programmed it?

All of it, most especially the audio code (no libraries there)... well bar a couple of handy little snippets supplied by Laurent De Soras.

And also analyzed the original hardware OSCar schematics, probed the actual synth and listened to the output (and various tapped points in the audio path) to develop the algorithms that were then coded.

Talk about handing someone their head...

LOFL.

[theblue1]

Originally Posted by audslu

You seem to lecture about some basic parts i know and not about what i don't know which makes us pretty close my friend, now matter how you want to speculate otherwise. It's just that i didn't understand your writing perfectly and engilsh is not my native language. About the error signal you confused me there, you don't have to analyse it into two signals there's no such thing, it's a single one both in analog or digital, don't create new definitions, others won't understand what you re saying.

What's with Fourier?? Obviously you don't understand digital audio. What do you think plugs and algorithms do, process time? How softsynths generate waveforms, they put samples in your screen to watch? You don't have a clue my friend and i was wrong to ask you it seems.

FFT is a vital part of digital image and the whole digital world and you think it's about time only (that's the small part), clearly you re wrong. It's how a cpu processes audio, what did you think it waits for the sample points to arrive and then alters the timing?!

What's with fourier? The proof to the almighty theory from which digital audio is being born, that's what it is with.


[...]

Originally Posted by audslu

Ok chill. With all due respect Mr Hodgson, but it seems we don't understand each other very well and i'm terribly sorry about my lack of the english languange terminology, but lets don't play with semantics. I don't claim to be an expert, i have already stated that. I like conversations about digital audio, and i might also learn something new from this. With you saying"Oversampling is not jitter" i thought "no kidding?" ... what is this guy talking about?

About oddity filter if you move the cutoff fast enough will you hear what i'm saying. Another example is pro53 filter, more noticeable there. Newer softsynths don't do that, and i'm not talking about diva only.

I was kind enough not to mention i don't like imposcar (1 or 2), never liked it, sorry. Pure coincidence or not, that's another topic.

I have heard many softsynths (including freeware) as good or even better made with synth-edit, it doesn't mean this guys are digital audio experts neither do they build converters or design converter chips! Programming a synth does prove digital audio expertise to me nowadays. Also there is freeware parts of code and opensource projects out there, so one can copy and learn code along the way.

Cheers.

bold added to highlight audslu's rudeness and insulting behavior

Extraordinary.

[Jon Hodgson]

Originally Posted by audslu

I was kind enough not to mention i don't like imposcar (1 or 2), never liked it, sorry. Pure coincidence or not, that's another topic.

That's fine, everyone is allowed their own taste.

I'll just have to settle for Sandy Vee, Rick Wakeman, Billie Currie, John Foxx, Jamiroguai, Underworld, Trent Reznor..... I could keep typing but I think you get the point....

I have heard many softsynths (including freeware) as good or even better made with synth-edit, it doesn't mean this guys are digital audio experts neither do they build converters or design converter chips! Programming a synth does prove digital audio expertise to me nowadays. Also there is freeware parts of code and opensource projects out there, so one can copy and learn code along the way.

Since I specifically said I designed the algorithms based on actual research into the original hardware synth (they're modifications of known algorithms in much the same way as the hardware filters are modifications of known filter architectures), and also that I didn't use any library code in the audio processing, what you've just done is either demonstrate you're not paying attention, or accused me of lying.

Anyway, it's not a language problem getting in between us, the problem is that you quite clearly have got a lot of things in your understanding of signals in general, and discrete time signals in particular, very VERY wrong.

[audslu]

@ blue I was wrong he has a clue. What's opinion have to do with insulting? Are you the forum politeness or kis-ass moderator? Well the world can be cruel and bold, something you brag about others don't even like sometimes. Life sucks.

@hodgson I was wrong you have a clue.

[theblue1]

Originally Posted by audslu

@ blue I was wrong he has a clue. What's opinion have to do with insulting? Are you the forum politeness or kis-ass moderator? Well the world can be cruel and bold, something you brag about others don't even like sometimes. Life sucks.

@hodgson I was wrong you have a clue.

No, but I thought your dismissals of him (highlighted above) -- when any half-educated fool should have been able to figure out he clearly had a good grip on what he was talking about -- demanded some form of comment. And then, after many of what must have been quite frustrating attempts to get through to you regarding your own knowledge gaps but being dismissed as "not having a clue," he, with seeming reluctantance, pointed to a virtual synth he coded, you went on to insult that... sheesh.

You should have heard what I was thinking.