44.1 vs 96khz music - double blind study conducted...

[Spe3D]

Well, if it helps here is a similar thread about sample rate from only the other day. This is what the working reality of how we all use our daws re. sample rate. (mainly at 44.1 kHz - with the need for higher on some occasions because of the noise or headroom etc)

http://www.kvraudio.com/forum/viewtopic.php?t=212918

[rifftrax]

it doesn't matter if you can't hear that particular frequency, it still effects clarity.

oh right...the famous "utter bullshit" theorem

good luck with your audio rocks and wave-propagation corrected MIT network arrays.

do you also use cables that correct for the inherent RLC artifacts that plague today's audio? a travesty that one is.

[afreshcupofjoe]

With digital audio, quality decreases as frequency increases. 22kHz @ 44.1 is a saw wave... i'm sure you know.. it's called the Nyquist criteria.

You have a very poor understanding of the Nyquist-Shannon Theorem. Go back to school. You do not understand what you are talking about.

Nyquist criterion is sampling at a rate two times the maximum signal frequency to sufficiently describe that frequency without ambiguity. Tell me how I'm wrong.

Strictly speaking what I said is true; however it does not take into account time averaging.

I have to sleep, so I'm going to say this as quickly as possible.


Your statement, "Quality decreases as frequency increases" implies that accuracy of digital sampling gets increasingly worse the higher you move up the frequency spectrum. This is not true. Nyquist shows that the audio will be perfectly represented digitally all the way up to the nyquist frequency, and then it is basically not represented at all. Anything higher than Nyquist frequency just creates aliasing. A frequency captured above nyquist is not "a saw wave" it just shows up as aliasing, which is completely different than a saw wave, and much more complicated. However, you will never even hear this aliasing, because it all gets filtered out below the nyquist frequency. Even if you had bat ears, you wouldn't hear anything above 22,000Hz because it just gets filtered out. It's not even there.

It's not like:

frequency-->
Good Audio... Good audio... Good audio... |Nyquist Freq| Worse Audio... Bad Audio... Horrible Audio... Noise

It's more like:

frequency-->
Perfect Audio... Perfect Audio... Perfect Audio... |Nyquist| Nothing....

So unless you can hear above 22,000 Hz, you are not missing anything. And you are not getting bad vibes from some kind of digital distortion happening in the inaudible frequency range, because that is not how digital audio works. There is nothing there.

I hope the explanation above helps some people understand Nyquist better. I didn't really get it at first either, and I remember it really making the difference when I realized that it only took 2 graphical points to represent a sine wave and that Nyquist was like a brick wall where those 2 points couldn't represent a sine wave of any higher frequency.

[debra1rlo]

it doesn't matter if you can't hear that particular frequency, it still effects clarity.

That's exactly what my dog told me. I'm glad I took him to the store with me to help pick out my new monitors.

[jupiter8]

Ok, a simple naive question as i'm not so knowledgeble in these matters. Assume you can somehow pick up on frequencies above 20 kHz and it affects clarity and you can measure the brainwaves changing,how come it does'nt show up on the blind tests ?

[Per Lichtman]

Regardless of whether we agree or disagree on this topic, there is no need to attack or name call. Thanks to everyone that has made an effort to maintain a professional tone on this apparently controversial issue.

An implication was made earlier that I had made frivolous statements about digital processing. I don't like dealing in maybes so when I have had questions about whether something made a difference, I have created many files and performed blind or double-blind testing. Sometimes things I thought would make a difference didn't and sometimes things I didn't think would make a difference did. I found that often setting up the tests took a great deal of time, care and effort and in my previous posts on this topic I have volunteered to share files to illustrate points and responded to feedback. If you doubt the validity of my tests, feel free to setup your own. My comments on this topic have been meant to help take existing tests further and to help people with different viewpoints to prove to each other what they are talking about. It was never my intention to attack or defend the test. If I have offended anybody in the process, I apologize.

Per, I'm not sure I understand what you're saying.

It seems, as has been discussed here earlier, that one way to test higher samplerate and bit depth differences in the final output that would rule out differences at other stages would be to create the file at 24/96 or 24/192 and then resample that file to Redbook specifications.

That is what they did here. The music was at 96khz to begin with and it was resampled to Redbook specifications, eg. 44.1khz.

it does not provide evidence that there is no benefit to working at higher samplerates during production but rather only that the final output can be oversampled without the listener losing anything.

How is any oversampling being done? Again, the source material is 96khz. They are DOWNsampling it to 44.1khz to see if anyone could tell the difference, and no one was reliably able to.

Maybe I'm just not understanding the terminology here.

There was discussion in the other forum that the particular testing of methodology used in this case introduced a potential unknown by adding an extra element to the signal path in the case of the converter as it was apparently a realtime part of the signal path. I was suggesting that one way to rule that variable would be to use a shorter signal path and create a lower sample rate conversion of the file ahead of time so that they could go between the two as some people were concerned that the converter was never allowing the signal to be distributed at full fidelity. I do not know whether this actually affected the results or not. I suggested a way to rule that out as a possibility.

I have to thank you Zircon: You caught a typo on my part. I had said "oversampled" when I meant "downsampled" in the segment you highlighted at the end.

Look at it this way. When you import files of any samplerate into a DAW project - the individual sample points are either going to be misaligned and require upsampling and conversion to achieve a downmixed output at 1 specified sample rate or they are going to be snapped to line up with each other in the case multiplicatively related samplerates (22.05, 44.1 88.2, 176.4, etc.). The smartest thing (duh) would be to avoid the conversion process altogether and record anything with sample rates that are multiples of each other so that conversion simply involves sample-point doubling. Herein - you've not created ANY new information AT ALL in the sampling or time-related domain. Period. End of story. I mean honestly - you are saying that by using a high-precision equalizer that oversamples I have created new and valuable sample data that 'should' be used as part of the math in future signal processing to ensure quality of the "mix"? Bullcrap. Equalizers and compressors and flangers and phasers don't operate to alter the y-axis extension of the audio. That's not the function of them (they are not pitch-shifters if I remember correctly) and if that's what is happening then I'm pretty sure they are not doing their job...are they?

This is math guys. Not science fiction. Keep your x and y axes (yes, that's the plural) separate and you'll do okay.

rifftrax, I would appreciate a more constructive tone in future responses as there was no need to for the use of implied obviousness or all caps section. I am not saying that you are creating new information in the waveform with higher sample rate processing. I am saying that digital processing involves limitations. Several equalizer designers, including George Massenberg, have talked about how there can be approximation and aliasing errors at lower sample rates. I have used several equalizers that do not oversample, had them process the same clip of audio at different sample rates (with several different methods of resampling used during testing to bring the audio from sessions at one sample rate to another) and found that in many cases, I could take a file that started out at 44.1khz, process it at 192khz, save that file, sample it back to 44.1 khz and find that the result was different (and in several cases preferable) to results of staying at 44.1khz all the way through. I am not suggesting that they are adding information but rather that one approach to audio editing might leave fewer artifacts in the process of making changes than another.

My last statement was confirmed with blind testing and (as with anything that produces a difference) the upsampling approach is not always going to be preferable or useful. There is, however, a real difference that I took time to demonstrate. With the help of other people on this forum I was able to, among other things, help the employees of Sonalksis to find a bug in the RTAS operation of their SV-517 EQ under the newest versions of Pro Tools LE by demonstrating it with empirical evidence. As a side note, it remains unclear whether the issue is in the plug-in itself or the host but that's another discussion entirely. If you would like to disprove what I have stated, I'd be willing to provide you with the assets used to conduct some of theses tests and links to freeware plug-ins that can be used for the purpose.

[deaf dunderkwac]

the important point is that the audio was produced at 96.

That's where the maths gets a bit easier for the sample errors to be corrected (jitter and whatnot).

The (imho @ 96kHz@24bit or higher) better-produced audio is then down-converted and the subjects have a hard time hearing differences and you all seem to be arguing about the differences in producing at different sample rates.

Interesting.

When I 'could' hear it was quite apparent that there was less jitter and other abnormalities at higher sample rate and deeper bit-depths compared to audio using 44.1@16 bit conversion. But.... (the big but) this was only really apparent during the production. I still think audio gets a remarkable increase in fidelity based on the number of bits per sample, rather than the sample rate of the conversion.
Audio produced with higher sample-rates and bit-depths was heard (by me) to have less 'issues' when down-converted
caveat-if the down-conversion of the original was at an exact even multiple of the original i.e 88.2 to 44.1 or 96 to 48. This may have changed with recent converters but I'll never know, as I'll never hear it


Good read/study anyways.

[ericj23]

With digital audio, quality decreases as frequency increases. 22kHz @ 44.1 is a saw wave... i'm sure you know.. it's called the Nyquist criteria.

You have a very poor understanding of the Nyquist-Shannon Theorem. Go back to school. You do not understand what you are talking about.

Nyquist criterion is sampling at a rate two times the maximum signal frequency to sufficiently describe that frequency without ambiguity. Tell me how I'm wrong.

Strictly speaking what I said is true; however it does not take into account time averaging.

I have to sleep, so I'm going to say this as quickly as possible.


Your statement, "Quality decreases as frequency increases" implies that accuracy of digital sampling gets increasingly worse the higher you move up the frequency spectrum. This is not true. Nyquist shows that the audio will be perfectly represented digitally all the way up to the nyquist frequency, and then it is basically not represented at all. Anything higher than Nyquist frequency just creates aliasing. A frequency captured above nyquist is not "a saw wave" it just shows up as aliasing, which is completely different than a saw wave, and much more complicated. However, you will never even hear this aliasing, because it all gets filtered out below the nyquist frequency. Even if you had bat ears, you wouldn't hear anything above 22,000Hz because it just gets filtered out. It's not even there.

It's not like:

frequency-->
Good Audio... Good audio... Good audio... |Nyquist Freq| Worse Audio... Bad Audio... Horrible Audio... Noise

It's more like:

frequency-->
Perfect Audio... Perfect Audio... Perfect Audio... |Nyquist| Nothing....

So unless you can hear above 22,000 Hz, you are not missing anything. And you are not getting bad vibes from some kind of digital distortion happening in the inaudible frequency range, because that is not how digital audio works. There is nothing there.

I hope the explanation above helps some people understand Nyquist better. I didn't really get it at first either, and I remember it really making the difference when I realized that it only took 2 graphical points to represent a sine wave and that Nyquist was like a brick wall where those 2 points couldn't represent a sine wave of any higher frequency.

just to clarify that your notion of perfect audio up to nynquist and then zip after wards just isn't true - or it requires certain conditions to be met

http://www.digital-recordings.com/publ/ ... onclusions

that shows the need for interpolation and the errors that can generate

and then we all know that your original signal must contain no information above nynquist - it must be bandlimited - alas the brick wall filters that seem ideal for this have a habit of generating ripple effects

To be clear I am not saying that the above sources of problems can't be overcome, just that they have to be - and I suspect that when cd's first arrived these things were not as well implemented as they are now

[sergef50]

Let me clear this up and draw a line on this and give you some facts....


End product for you of for the consumer that:

This test proves thats for end product there is no scientifically provable reason to run at doublerates....FACT FACT FACT!!!

=-------------------------------------------------------------------<THE LINE

Processing during DSP for an engineer:

There are benefits of less aliasing and more headroom to be had by increased BITRATE and SAMPLERATE during recording and processing...great benefits...should you wish or be able to use them.

END OF STORY.

[Sascha Franck]

I have used several equalizers that do not oversample, had them process the same clip of audio at different sample rates (with several different methods of resampling used during testing to bring the audio from sessions at one sample rate to another) and found that in many cases, I could take a file that started out at 44.1khz, process it at 192khz, save that file, sample it back to 44.1 khz and find that the result was different (and in several cases preferable) to results of staying at 44.1khz all the way through. I am not suggesting that they are adding information but rather that one approach to audio editing might leave fewer artifacts in the process of making changes than another.

You have a point here and I thought nobody would argue about this, especially here at KVR where aliasing debates are like a pet peeve.

I have made the same sorts of tests and sometimes the results were just dramatically different. Even more so once whatever softsamplers were involved. And yet even more so in case I used proper up/downsampling algorhythms instead of whatever comes with your host of choice (such as to be found in the free, x-platform Audiomove or Voxengo's RBrain).

As said, I thought it was common knowledge that working with higher sample rates (note: I said "working with", not "delivering in") can in many situations improve the overall sound quality.
Admittedly, the situation isn't as bad as it has once been, with plugins (or whatever digital "devices") using proper internal oversampling algorhythms and what not, but it's still unarguably existing.
Personally, I still can't justify the huge extra efforts it'd take to work in 96kHz (constant freezing/bouncing, plus a fair bit of upsampling work, should some sound sources come from CD, older productions or whatever) and fortunately my "productions" (if I may even call them that) don't seem to suffer from the issue too much (of course, as a scientifically interested person, I made some tests for myself).

_______________________________________

Anyways, I do however have a question for those in the knows. I never exactly managed to test things myself and I haven't exactly found an explanation yet either. Could even be that aliasing is related, but if so, I probably haven't fully understood the "concept" behind aliasing (which pretty much likely is the case, I've always been more of a "hands on" kind of person).

Whatever, here goes:
Those of you playing guitar (like me) might have noticed the "phenomenon" already as well. When you play, say, a rather high small interval, such as a minor second somewhere on the high E and B strings, you will sometimes notice certain deep notes to occur. With clean sounds, this usually gets unnoticed, but with higher amounts of distortion, the effect can sometimes become quite drastic.
In lack of a better term I'd call this "subtractive harmonics". I'm not saying that the lower frequency coming up is exactly A - B, but obviously that would make sense. Also, the effect is more noticeable the smaller the interval gets.
Along similar lines: You may know the method of tuning a guitar by harmonics. Such as playing the natural harmonics of the (low) E string at the 5th fret and of the A string at the 7th fret simultaneously. What you will listen to to achieve the best tuning is a sort of "oscillation" going on. The closer the two harmonics get, the higher the "oscillation rate" gets. In case the two are not exactly close, the rate of that oscillation surely would be in inaudible range (just as most LFOs). But once the pitches get closer, the oscillation rate is getting faster, and it may as well become an audible rate (just that at this point in time, the volume is already decreasing a lot).

Please note: The two examples to me are describing different problems. In the last "tuning example" a difference of a minor second (as used in the first example) would usually result in a too low oscillation to be audible at all, at the same time whatever audible low interval (as mentioned in the first example) might occur. Could be related to interferences between harmonics or whatever.

Ok. So much about my observations.
What seems to happen is that if two frequencies are quite close together, there's gonna be a quite lower frequency occuring as the result.
Now, if this is happening using source frequencies in the audible range, why shouldn't it as well happen with frequencies that are outside the audible range? I couldn't see a reason why it wouldn't happen as the laws of whatever physics should pretty much not be limited by whatever human hearing ranges.
In other words: IMO it could very well be possible that higher frequencies (even way outside the audible range) close to each other would produce "substractive harmonics" that'd be very well inside the audible range. They might be too low to be perceivable. Or not. I wouldn't happen to know.

Anyways, this might be one of the reasons, why working (still not "delivering") at higher sample rates may have it's benefits.

Does anybody know whether such a phenomen happening in the inaudible range does have an effect for the audible range?
And: Is what I'm describing here related to aliasing? I always only see aliasing in relation to plugins and what not, but what I've been trying to describe is clearly happening without using any plugins at all, just two recorded single frequencies will do to cause the effect.

Hm, I could probably do a test myself, just generating two frequencies outside hearing range and then playing them along each other...

[nollock]

Whatever, here goes:
Those of you playing guitar (like me) might have noticed the "phenomenon" already as well. When you play, say, a rather high small interval, such as a minor second somewhere on the high E and B strings, you will sometimes notice certain deep notes to occur. With clean sounds, this usually gets unnoticed, but with higher amounts of distortion, the effect can sometimes become quite drastic.

http://en.wikipedia.org/wiki/Intermodulation

[nollock]

Does anybody know whether such a phenomen happening in the inaudible range does have an effect for the audible range?

Certainly inaudible frequencys above 20kz can be processed and result in frequencys bellow 20khz. Ring modulation creates sum and difference frequencys, so ring modding 5khz, with 21khz, would result in 16khz, and 26khz, the former being audible.

The same can happen with a non linear process like distorition, IE intermodulation distortion.

That said in my experience it doesnt usualy sound good, it usualy sound effing awful.

[nucleon]

Not sure if its been posted already but heres an insight into why some may percieve a difference, from one of Lexicons audio physicists.

A form of speaker distortion when playing higher samplerate material.

Its the power point presentation called "Slides from the AES convention in Banff on intermodulation distortion in loudspeakers and its relationship to "high definition" audio."

With the intro: "And now for something completely different... Being currently over 60, and having in my youth studied information theory, I have a low tolerance for claims that "high definition" recording is anything but a marketing gimmick. I keep, like the Great Randi, trying to find a way to prove it. Well, I got the idea that maybe some of the presumably positive results on the audibility of frequencies above 18000Hz were due to intermodulation distortion, that would covert energy in the ultrasonic range into sonic frequencies. So I started measuring loudspeakers for distortion of different types - and looking at the HF content of current disks. The result is the paper below, which is a HOOT! Anytime you want a good laugh, take a read."

http://world.std.com/~griesngr/

http://world.std.com/~griesngr/intermod.ppt

(Not to be confused with the oversampling of synths to avoid aliasing, which should be done internally. Recently found even some NI plugins render differently at higher samplerates, and appear not to have an oversampling mode to recreate this at lower rates, like other plugins).

[Sascha Franck]

Thanks for the link, nollock. A bit too mathematical for my brain right now, but certainly explaining things properly. Will try to understand what exactly is happening.

That said in my experience it doesnt usualy sound good, it usualy sound effing awful.

Well, I wasn't saying it sounds good at all. In case this occurs on some guitar vocings, I usually always try to either use another (mostly less distorted) sound or find some other voicing.
Yet, it's just part of whatever we're dealing with almost all the time. For instance, have two vocalists sing a melody doubled and the effect will always be present (at whatever perceivable or non-perceivable volume) as they will never be 100% in tune with each other.
In other words: It just belongs into music. Whether it's audible, whether it's "nice", whatever - that's pretty much beyond the point. I was just thinking about possible side effects that using higher sample rates might have.

Oh, and I must look up the secrets behind ring modulation as well. Always wanted to know, so now could be a good point in time.

Thanks again.

[Rangtangtang]

The equipment list included amplifiers from high-end manufacturers like Adcom, Carver, Sim Audio and Stage Accompany, and speakers from Snell and Bag End, as well as the oft-worshipped Quad ESL-989 electrostatics, which are supposed to have usable response up to 23 kHz — which is, of course, above the Nyquist frequency of the HHB recorder's converters. The subjects listened to discs that covered a wide range of material and included classical instrumental, choral, jazz, rock and pop, from audiophile labels like Mobile Fidelity, Telarc and Chesky.

For a proper test the tracks would have had to be recorded at the 96khz and 44khz rates in the first place, not converted to? Of course whith the best playback system 44khz should sound at its best. I find that there is less listening fatigue at higher rates.