I recently met a musician who claims he can quite easily hear the difference between 44.1KHz and 96KHz.

This shocked me a little because I'd always been told that the human ear cannot hear any higher quality than CD (44.1KHz) quality.

So... was this guy just lying (or fooled by his senses), or was I being lied to when I was told the human ear cannot hear any higher quality than CD?

The most profound differences are not higher frequency response but the effect of the anti-aliasing filters on the frequencies that can be heard. At 96kHz sampling rate the filter effects at 20kHz are somewhat different than when sampling at 44.1kHz. This is a measurable result but whether or not many, or any, people can truly hear it does not seem to have been established by impartial studies. Or maybe it has but the results don't please the people who promote higher sampling frequencies.

The results of those filters are not exactly a given anyway. There are a variety of ways, and many graduations of these, to accomplish the filtering, with different final effects. There are those who claim to prefer digital recording and playback with no such filters, regardless of the images. There are some who have developed special (non-conventual) processes to produce CD standard files from high sample rate masters without anti-aliasing filters, but also without the stronger images that would result from the more normal downsampling minus the filters. Most or all of this is too subtle for anyone to really hear it if they don't know to which form they are listening at the moment.

There are some limited studies that indicate people may be able to respond to frequencies well above 20kHz (maybe as high as 30-35kHz) under very limited conditions. The mechanisms seem to be something other than hearing, such as direct conduction of vibrations through parts of the body. Such high frequencies are very rarely strong enough to accomplish this.

call him/her on it... make em prove it to you

it is possible that they can hear a difference, but i think it is highly unlikely

record something at 96khz... and make a copy at 44.1khz (resample with something good like SSRC) keep the bitdepth the same

Andy, I didn't realise antialiasing filters were applied to music. Do all audio players use antialiasing?

If no one can hear higher than 44.1KHz, what is antialiasing needed for? Surely it would be too fast to tell the difference between the original aliased sound and the smoother antialiased sound?

MOST audio players (as part of the DAC) use anti-alaising filters. The image is reflected back down from the Nyquist limit. That means it gets mixed into the music. You generally can't detect it as something separate, on its own, it just adds stuff that should not be there.

It comes in reverse order. The lower the signal frequency, above the Nyquist limit, the higher the frequency of its image. Which also means, the higher the frequency above the Nyquist limit, the lower the frequency of its image.

Nyquist limit = 22,050 Hz at 44.1KHz sampling rate
image of audio at 24kHz appears at
(22,050 Hz - (24,000 - 22,050) = 1950Hz) = 20,100 Hz
image of audio at 28kHz appears at
(22,050 Hz - (28,000 - 22,050) = 5950Hz) = 16,100 Hz

Duhh... sorry, that all went completely over my head!

I know that oversampling in A/D and D/A converters are employed to both spread the same amount of noise over a wider spectrum (noise shaping) and to avoid aliasing, even with 44.1 khz.

Digital Finite Impulse Responce filters can be used to filter the signal.

Enrico

It's not half as hard to understand as it sounds.

the Nyquist-thing:

you need double the frequency (samples per second here) to record a certain frequency (frequency of the signal wave)

If you want to record a Signal that's 800 Hz you need 1600 Hz (practicaly a bit more afaik, but this is theoretical) to record it

If you'd record a Signal of 800 Hz with a samlingrate of 1500 Hz (1500/2 = 750) there would be 50 Hz too much, those "come back down" and appear at 750 Hz in this case. This is a not really recorded signal and unwanted, the so-called Alias-signal.



So if I'm right a Anti-alias Filter is a low-pass filter (lets signals below a certain frequency go through) that cuts off every frequenzy above half the sample frequency to avoid alias-signals.

The Problem is that those filters can't be "perfect" but if you apply them at a higher sampling-frequency the result will be closer to perfect and might sound a tiny bit better, but that's unaudible for nearly everyone.

hope I didn't tell something wrong and also hope that it's a bit clearer now,
and sorry for my bad english

Oh I think I get it.

It's a lot like picture resolution. In order to resolve a certain level of detail, you need ideally more than double the number of lines than the intended resolution. Ideally as many as possible. If your resolution is too low, the detail blurs together to form new colours that were never part of the original image.

right, it's principialy the same

I found a nice explaination somewhere a few days ago, but I think there's no need for searching since you know how it works.

The main reasons for the higher rate are quite good explained (I think) in AndyH-has post below your original question.

simply put:
frequencies so high that you can't hear them, produce digital conversion artifacts in the range you *can* hear. This phenomenon is called 'aliasing'.

Antialiasing filters block out those artifacts.

Greetings !

I was just Googling about this right now. There's a site that says that anti-aliasing filters are already good enough at 44 KHz. You shouldn't probably hear the difference because of filter problems. The matter seems to be simpler than that. For example, take a violin and a cello. They may produce faint harmonics at ~ 30 KHz (let's say they the former has one higher peak at 30 KHz and the latter, at 32 KHz). If you listen to them (that is, nothing to do with recording), you may hear a 2KHz beating. However, when you record them, you would do that separately. Record the violin at 44 KHz and you'll lose that important peak. Then record the cello and you'll lose that other important peak. Now mix them together: there's no 2 KHz beating ! If you recorded them together you could sample at 44 KHz and still get the beating. Since you don't, then you'll have to record at least at ~65 KHz. 96 KHz would then be a convenience sampling rate (ie 2x48 KHz).

That's what I read. I'm no speciallist on that. You may google for it also.

Henrique Dante de Almeida

Why may I hear something like that ?

Sebi

Hi,

this is a 2Khz (stereo) square wave, represented in 16/44.1 PCM



A square wave is actually composed of a sine-wave fundamental (of 2KHz in this case) with an infinite number of it's odd order harmonics folded back into it (3rd, 5th, 7th etc). In fact a'perfect' squarewave doesn't exist, it would have an infintely short rise and decay for each cycle, requiring an infinite number of harmonics, but the more (higher-frequency) of those odd-orders you add, the closer you get to one. This is how the 'edges' needed for digital data transmission are created on such things as analogue phone lines.

This waveform obviously doesn't exist in 'nature', there's no way of producing it acoustically, transmitting it through the air and capturing it with a microphone, it has to be synthesized.

So, this sythesized 2KHz sqaurewave actually has harmonic components extending to 100's of KHz and beyond. Strange but true. You can't 'hear' them, but they're there, they create theis waveform by reinforcing or attenuating the original 2KHz sine.

To actually reproduce this wave 'perfectly' in the analogue domain as the output of a DAC (that is, downstream of it's anti-aliasing filter) is as 'impossible' as the waveform itself is. Filter ringing and phase-shifting between frequencies will produce various effects such as rippling which can be seen graphically if the output is re-captured digitally or monitored in real-time on an oscilloscope.

Now as it happens almost *all* musical instruments produce sound swith harmonic components extending to 40KHz, 50KHz and beyond. Some, such as muted brass produce very substantial pressure levels indeed at these frequencies.

Can we hear them, or sense them in any way? Doubtful (even if you go with the putative non-aural mechanisms some suggest).

BUT they are nonetheless intrinsic to the waveform which results when they are captured - it is *irrelevent* that we cannot 'hear' them, or that the recording hardware or digital protocol is 'band-limited'.

On playback of a recording, the same digital-filtering effects which can be seen graphically in the output of the simple, mathematical square-wave will affect the ultrasonic components of musical instruments and *will* at the very least have an effect on timbre, from innocuous to possibly ear-shredding.

Please don't anybody tell me they *havn't* at some heard point heard a recording of violin or trumpet playing on a CD-based system that didn't make them want to clap their hands over their ears!

I'm not at all surprised to hear that a musician says he/she can hear their instrument reproduced more faithfully with higher sampling rate PCM.

Higher sampling rate = much more benign filtering and more realistic music.

R.

>>edits - yptos as usual.

OK., so far so good, though whenever someone brings up square waves in a discussion of digital, I expect the worst.



Wrong. If we can't hear them -- or their effects in the audible range -- then they are indeed irrelevant to our audio experience.



At the very most, that is *possible*, but not *certain* to happen, nor is it at all certain that whatever effect you hear on timbre you hear, is due to the sampling rate. You'd have to rule out lots of other causes. Generally the biggest 'hit' the accuracy of a digital recording takes is when the signal passes through the mic and the speakers -- the electromechanical parts of the chain. These are by far the least linear.




Please tell me that that you don't consider this proof that Redbook standard *necessarily* affects the timbre of a recording. (I've heard all-analog recordings that make me want to cover my ears, btw.)



I'm not surprised that pepoel claim all sorts of things. I'm far more surprised when they've actually tested those claims properly. Because that;s so very rare.





It can mean that. Doesn't necessarily mean that. It's down to how well the filtering is implemented.

.

It was just an example. If you were talking that the 2 KHz was a mistake, I'm sorry, it should be 1 KHz. If not, it's because of the following. I supposed that there would be an instrument which would produce a significant harmonic at 30 KHz (actually this is true, for example, for violins and flutes), and another that would produce it at 32 KHz. Since those frequencies are actually a pressure in the same medium (that is the air and then your ear), the expansions and compressions generated by the instruments will add to each other some times and cancel each other some other times at a rate of 1 KHz. Mathematically, cos(32KHz)+cos(30KHz) = 2*cos(31KHz)*cos(1KHz). Add the time in equation and you have a 1 KHz harmonic with a variable intensity of 2*cos(31KHz*t). In practice, you should have a few harmonics for every instrument in this region. For most of them, they will be so faint, that you won't ever notice it. The already cited instruments, however, are known to cause audible beating which enrich the listening experience. Unfortunately I have no link to that, except one that also claims this is true, but doesn't cite sources either :-/.

http://www.dvdsoftwareguide.com/all-about-dvd-4-guide.html

One should hope, thus, that every recording is made at very high sampling rates. After they are mixed and filtered with high quality equipment, they may be safely downsampled to human listening limits again.

Again, I don't think that filtering artifacts are relevant. The issues only happen with bugged filtering. Recent equipment shouldn't cause audible artifacts. Concerning the bad violin or trumpet, see the other discussion.

http://www.digitalprosound.com/Htm/SoapBox/soap2_Apogee.htm

Henrique Dante de Almeida

@gangran dizzy

i hear alot of audiophiles around me claiming alot.

They do have a hard time proving it... atctually they have failed all there claims when doing probally scientific correct testing.

I don't want to get into yet another nit-picking session on this.

You simply havn't grasped the point I'm making.

If the intrinsic ultrasonic content of our squarewave (of whatever frequency) cannot be maintained with 'time-domain coherency' on playback (within a band-limited playback-system), then QED - neither can any other *captured* sound/waveform which is defined by it's ultrasonic components.

The timbre of many, if not most, musical instruments *is* defined by this ultrasonic content. I'm amazed how few people haven woken up to this.

As I said - the ability to perceive/hear ultrasonic frequencies or to capture them discreetly with recording equipment is completely *irrelevent*.

Time-domain coherency is the key to realistic reproduction of music - instruments, voices, whatever.

Of course, if we're lsitening to a Stratocaster and an overdriven Marshall stack, this might all be moot.

R.

By reading your statement(s) here, one would think you have not paid any attention to the last several discussions on hi-rez audio on hydrogenaudio.org. Please go back and reference these prior discussions.

-Chris

It really isn't a matter of opinion.

Oversampling filters serve as a panacea for the limited resolution of RB CD (16/44 PCM).

But many people are building completely filterless/non-oversampling DACS. Why, one should be bound to ask?

Here's the rub; OS DACs do sine waves pretty well up to (insert freq; 10KHz?) but are utterly incapable of resolving a squarewave at anything close to this freq.

Non-OS DACs make an unholy mess of sines above 10KHz, but (at least some of them) can do squares at this frequency and beyond.

In a previous discsussion here at HA, someone siad that the the Non-OS DAC's inability to reproduce HF sines meant they were "broken".

Why, then, does the OS DAC's inability to reproduce HF squares not mean they are *edit >* NOT "broken"?

R.

I do actually pay attention, and not just to HA.

Do you have a point?

R.

I have grasped your point, as I'm sure the other post has as well: You think that ultrasonic information has some relevance to the audible waveform, though such has not ever been shown in a credible peer-reviewed study. But your premise as presented here is in error. 1. Assuming that analog filters are used for the anti-alias filters, no one has shown the phase distortion that occurs as a result to be audible. 2. Most systems today should be using linear phase digital filtes, in which pase distortion is not even an issue. 3. Since you can only hear frequencies <X, the waveform content >X is not relevant to audibility. You can not hear square waves, for example. You can only hear the sine waves composing the square waves which are <X frequency. The content >X is only relevant to a pretty looking graphical respresentation of the waveform, not to audible parameters. How you think supersonic content has to do with time domain coherancy of the audible waveforms, I have yet to figure out.

-Chris

I take that back. It might well be just as important.

That's really interesting actually. I just created a 96khz wav file with a 30k and a 32k tone and could hear a beat frequency as you say. Though at 2khz not 1khz

Why do you say the hi-res mix may be safely downsampled to 'human hearing limits'? If I resampled the above wav file to 44.1khz I ended up with silence!

Wrong. Resolution is a function of word length (bit depth), not sampling rate.
Oversampling filters were a solution to the difficulty (not impossibility) in implementing excellent filtering at 44.1.




Because there's no silly idea that some audiophile won't embrace. There are belt-driven CD players out there too. And no, it's not *many* people doing this. It's a relatively tiny cult, as with most audiophile tweaks. One question to ask is *who* is doing it.


I looked up your posts here and I see you were arguing for cryogenic treatment of wires elsewhere.


I suggest interested parties check out Nika Aldrich's comments on quare waves,
before they disappear from the the Web (they're now cache-only). Or buy his book on 'Digital Audio Explained' -

http://72.14.207.104/search?q=cache:stHyib...us&ct=clnk&cd=1

Listen to the re-mastered release of Steely Dan's "Can't Buy A Thrill". I gather it was transferred from 50KHZ (!) open-reel digital transfers of the analogue 2-tracks that Roger Nichols and the band saw as imperetive back in the early 80's.

From the get-go, there is an obvious 'beat' or pumping effect happening on the intro to "Do It Again" which renders it practically unlistenable. Probably a result of artifacts resulting from sample-rate conversion, reinforcing inate problems with PCM.

I'm flattered.

R.

You bore me.

The last word is yours, please do savour it.

R.

Hello,

For the few things that I've read here/know about SP, not using filters/aliasing reduction would create a violin++, instead of a normal violin. That may or may not be what the composer intended. Is this correct ?

For the square waves, human beings can't hear them. Our ears would filter them anyway. If the sound device already does this, then great, less work for my psychoacoustic organs.

Try it with two separate tracks (remember, that was the original proposal). Mix them together to produce the beat frequency. Now you have captured it in a file. Before you were only producing it as you listened to it. Now when you downsample, it can't go away.

Wow. You 'gather' that? Roger Nichols has a forum online. Maybe he could evaluate your theory authoritatively.

http://www.ioforums.net/forums/

I believe his current forum name is 'ioforums33'

No, I did mix it down to a file. As a quick way to do it I put a 30khz tone in the left channel of a stereo wav file and a 32khz tone in the right then mixed it down to mono. Play this back and I can hear a quiet 2khz tone. The frequency analyser only shows peaks at 30khz and 32khz despite me clearly being able to hear a sound.

I downsample to 44.1khz and the resulting file is totally silent. I can put flac files on the web to demonstrate if you want (obviously, you need a soundcard that supports 96khz playback).

Could it be that it's not really safe to downsample the sound ? :-D. Try downsampling to integer frequencies (ex: 48 KHz) to see if it works :-/

Henrique Dante de Almeida

http://www.ioforums.net/forums/view_topic....rum_id=1&page=1

If the beat frequency is an actual artifact that existed before recording, or as an audible artifact in the recording, it would be recorded within a 44.1kHz sample rate, since it would be an artifact in the audible range. However, if it is an artifact that exists after the fact(as some hardware will distort with intermodulation effects when presented with multiple high amplitude ultrasonic signals), then of course it would disappear when downsampled, since you removed the ultrasonic compoents causing the distortion(s). This was a sythetically created file, remember, not a recording of an actual sonic event that contains the lower band artifact distortions originally. It sounds like you may have hardware that is not well suited for ultrasonic content, unless you are just cranking the volume up to insane levels, exaggerating the problem that would not normally exist. However, it would not normally exist anyways, since the ultrasonic content would be tens of dBs under the main bandwidth midrange and bass levels.

-Chris

Same result if I downsample to 48khz. Though there is a beat-frequency at 2khz it is made up of two component continual sinewaves at 30khz and 32khz, PCM only considers the component elements not the result, in the same way as a square-wave is approximated through a large number of sinewaves. Sample below 64khz and the 2khz baby gets thrown out with the 30 and 32khz bathwater.

This is my (badly worded) explaination anyway, from my understanding of the Nyquist theorem...

Here is the 96khz file containing the two tones, I've doubled the mono up to stereo, please be careful with it.
Play this loud and you may very well DESTROY your speakers, headphones or amplifier
http://www.arafel.org.uk/~mandel/ha/ultratones.flac

Edit: Added a bigger warning

I did not use your samples, but I did create a file exactly as you described. The 2kHz byproduct tone is present in any loop back recording, and will survive a 44.1Khz sample rate file, since the 2kHz is a byproduct of intermodulation of the signals that must occur in the hardware analog stage. The synthetic file you created has nothing to do with a natural occurance. Being a digital creation program, it can create aritifical circumstances(such as seen here) where you can have two discrete tones exist without intermodulation behaviour(that must happen in the analog realm). But when this happend in the real world(analog realm), the intermodulation artifacts will be present, and will be recordable directly.

-Chris

I agree with all your points. However, to quote an article linked to previously in this thread (http://www.digitalprosound.com/Htm/SoapBox/soap2_Apogee.htm):



Here when mixing the 48 track recording down to stereo all in the digital domain there is a similar situation to the artificial one I created. Suppose that the first violins are playing a note with an overtone at 30khz and the second violins a note with a tone at 32khz, the close miking prevents the intermodulation distortion in the air from being recorded, however if all mixing is done at 96khz the distortion can reappear.

While as you say a loop-back recording at 44.1khz will capture the sound fine a straight downsample in the digital domain won't. The latter is what occurs when a CD is produced from a hi-res master.

Its a very interesting sound. My own analysis indicates there is no virtualy apparent 2khz sinusoidal content at all. But the beating of the high frequencies occurs in 2 khz non sinusoidal pulses.
The content of the beating is still ultrasonic though. I have a notion that this beating wouldn't be detected by the human ear because it depends on an ability of the ear to be affected (some sensory part to resonate) with 30Khz tone.
A pcm recording device could be affected by the tones at lower sampling frequencies because of the aliasing characterist of taking discreet level measurements rather than level integrals .
Then the representation of them is further complicated by their beating.
What would really reach the microphone, would be silence alternating with a 30khz (31?) tone, 2000 times a second, what is recorded is a variation in instantaneous level, heard in playback as an inaccurate reconstruction of that variations most likely cause.

I hazard audio should not be sampled at rates below the capability of mics to capture instantaneous level, or there is no way to do the antialiasing. The audability of the two high frequency tones should be a technical artifact of the sampling process, corrected by high quality downsampling. The experiment may highlight a problem with playback of high sample rate files.

regards'

Hey,

Could you try again, but with 30 KHz and 16 KHz tones ? Record them at 96 KHz, then downsample them to 48 KHz. (I am wondering if we cannot really listen to the 30+32 tones)

If I do that then all I can hear at 96 or 48khz is the 16khz tone making my ears bleed
Oh except that it sounds more of a clear tone when downsampled to 48khz. There is something removed by the downsample

This thread is fascinating! But most of it is too techy for me unfortunately.

Could anyone try again to explain how these byproduct tones/beats come about? And what is the low pass filter's involvement in this? (What is a low pass filter anyway?)

Well, here is what little I can remember from RF theory class (applies here as well). When you have a signal, say at 1kHz, and you send it through a non-linear element, such as the real world, you start producing harmonics at integer multiples, example at 2kHz, 3kHz, etc. Each of these is weaker than the one before though.

When you have 2 signals (sines, whatever) you still get the harmonics, and you start getting the cross-products, which is the beat frequencies you guys hear. Thing is, you only hear them cause they are getting sent through a whole bunch of non-linear elements, such as mixers, amps, speakers, the air, etc. BTW, sometimes you want very linear components in the signal path, for example power amps, etc. Sometimes you want very non-linear components, such as mixers, which are used in every single RF device you own, to bring in the stratospheric 2GHz or whatever frequencies down to more manageable ranges.

So, in theory, if you keep the 30k and 32k signals completly in the "digital/mathematical" domain, and you resample down to below their nyquist freq, and the resampling is "linear", then the signals should just disappear. If the resampling is non-linear, which could be on purpose to get a better sounding sound or whatever, then you might still get the 2k cross-product. In a way, the beat frequency is an illusion created by the real world.

Imagine two people walking side by side but with different length of steps. They can start pacing together on the same foot, but soon they will be out of step, and on different foots, some more steps later theyll get into pace again and some steps more -out of pace...the cycle will continue.
That cycle is the beating of the two different paces.
The PCM record is like noting down the gross foot position every click of a timer, so on clicks where both walkers feet are agreeing the gross foot position will be complimentry , when the walkers are directly out of step, each foot's position will counter the other and if directly out of step (and the feet are equal weight) their positions impression in the gross feet record will be zero.
In the record it looks like there is one foot moving at a similar rate but twice as much as each of the real feet, but coming in and out of existence in time with the beat period.

The question of how this beat period is generating an actual tone (other than the two tones related to the walkers paces) on playback of the record, is in contention in this thread.
Some believe that the beat period will be heard as a tone (walkers pace), or that the period will produced a tone in the air ~maybe a fluid dynamic phenomenon or something. It should be remembered that the beat isnt a tone though, its the periodic rising and falling in loudness of a tone.

The frequency of the tones being examined here is ultrasonic but playing their PCM record is producing an audible tone with the frequency of the beat period. This is suggesting to some people that beat periods are hearable as tones.
A lowpass filter cleans a pcm record of all mathematicaly apparent suggestions of sinusoidal oscillations above a certain frequency, (allows all the lower frequencies to pass).
When the ultrasonic test signal is lowpassed to remove theoreticaly inaudible frequencies from the record, the 'sound' of the beat frequency disappears as well.

The question is, is the 'sound' of the beat frequency something we would hear in real life or is it an error of the digital recording and playback process?

I guess the best way to proof that it has influence on the realworld-sound is to produce such a signal in the realworld. If we can't it's not a proof that it's impossible, but if we can...

What would be needed to do so?

The sound of the beat freq has nothing to do with the digital process. You could say it is a completly analog phenomenon. It will happen with a digital source as well as with some high-bandwidth analog tape or a vinyl deck or whatever. It is just the mix of two different signals.

Did you note the relative level of the intermodulated product? It was only a few dBs above the noisefloor, when I generated the 30kHz and 32kHz tones at *-3dBFs! Even then, it was only audible when I cranked the volume high enough that would never be usable for music listening. This is totally unrealistic circumstance. First of all, the difference tone would be masked in real music due to the other signals. Second, the relative levels of harmonics in ultrasonic range in real music is tens of dBs under 0 dBFs, not -3dBFs. The effect as seen in your example, for the most part in real circumstances, would be buried under the noisefloor, and if it did manage to occur over the noisefloor, it would likely be masked by anything in the lower bands.

-Chris

* Edit notice: The final amplitude of the waveform after generating and mixing these individual waveforms was -3dBFs. It appears that my poorly worded statement could be read as if I mean that each individual waveform was -3dBFs before mixing.

So you added two sine waves with an amplitude of -3dBFs? In that case the result will clip and that may explain the tone you heard..

I know a beat will happen whatever, but a sound is generated in some playback configurations and not others.

No ones been listening to clipped test signals

It has been mentioned that in the real world very high frequencies like this may produce audible products, but with these test signals no post processing has been done to the sine waves to simulate possible real world phenomenon, the impression of two ultrasonic sines should be all that is present in the PCM record -no subtle air dynamic phenomon in this signal. But an unexplained noise is present on playback. Since, high quality lowpassing leaves no audible noise, it is consistent with the virtual sines not having generated any extra audible product.
So threre really appears to me to be a problem in the playback. Maybe this is lax internal resampling in the soundcard, or patterned rounding error, i can only guess.
Maybe try same experiment at 24 bits, or try playback through a ~XiFi