Skip to main content

## Topic: DSD vs PCM, bit vs hz (Read 26585 times)previous topic - next topic

0 Members and 1 Guest are viewing this topic.
• saratoga
DSD vs PCM, bit vs hz
##### Reply #50 – 16 February, 2014, 04:04:50 PM
It might be, but again, 8/9 is a relatively simple ratio

This is irrelevant.  Each of those tones has a finite bandwidth because of the transform limit.  Say each one has a width of 0.5Hz.  They would also include ratios such as 800.1/899.9 = 8001/8999.  This is not a simple ratio.  As you can see, it makes no difference.

Preferably mutually prime numbers.

8999 is a prime number.

You should get a decent hash, but still not high

Indeed, <-129 dB is not very high.

DSD vs PCM, bit vs hz
##### Reply #51 – 16 February, 2014, 09:15:05 PM
Arnold: Your graph is worthless, because you used (mostly) harmonically related frequencies. That doesn't work for IMD testing.

Anybody who can accurately read the frequency scale of that chart knows that the above stattement is incorrect.

The frequencies were arbitrarily chosen - 20, 25, 30, 40, 50, 60, 70, 80. 90. 100, 150, 200, 250, 300 and so on. The last octave is 10 KHz, 11 KHz, 12 KHz, 13 KHz. 15 KHz etc.  a few are harmonic, most are not.

Please cite an authoritative independent source that confirms that exceptional claim. I apologize for sending you on a wild goose chase because the physical law that you cite simply doesn't exist.

Quote
Please check how CCIF IMD is done and how SMPTE IMD is done. The frequencies used in those do not have simple common divisors.
(E.g. CCIF: 19/20. SMPTE: 30/7.) Essentially your test is CCIF times 10. It is pretty hard, but only when tried in top octave. At low frequencies such simple ratios are generally handled well and SMPTE is harder. Then,

If you could cite a credible authority, I would rerun the test with those test tones. I know for sure that the physical law you claim simply does not exist because I've run a wide variety of 2 tone, 3 tone and multitone tests on a wide variety of equipment types and models. I've also done the math.

Here is an example of Agilent, a highly respected test equipment manufacturer recommending the use of harmonic frequencies in their technical literature:

http://cp.literature.agilent.com/litweb/pdf/5991-2396EN.pdf  (Figure 1)

One reasonable guide for picking  multitone frequencies is to pick frequencies that line up with the FFT bins so that a multitone with flat frequency response is more likely to present itself that way when analyzed by a FFT.

Please check that graph again. The frequencies in that particular multitone are arbitrarily chosen - 15, 20, 25, 30, 40, 50, 60, 70, 80. 90. 100, 150, 200, 250, 300 and so on. The last octave is 10 KHz, 11 KHz, 12 KHz, 13 KHz. etc.  a few happen to be harmonic, but most are not.

Multitones of any kind are easy enough to create - another possible sequence of frequencies are the standard 1/3 octave frequencies. Doesn't matter. Another possible sequence involves the frequencies that divide into the sample rate with integer quotients and no remainders.  Audio Precision recommends a logrithmic sequence. For IM testing it doesn't matter that  much.

The ratio of the test frequencies does not matter within reason. The spurious responses created depend on the orders present in the nonlinearity, but they are always created. If you use harmonic tones some spurious responses due to harmonic distortion land on top of spurious responses due to IM distortion, but there are plenty of circumstances where that does not happen.

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

http://www.prismsound.com/test_measure/sup.../multi-tone.php

http://www.klippel.de/measurements/nonline...distortion.html  (extra good references cited)

http://www.ap.com/kb/show/60

Higher frequencies tend to create more IM products because amplifiers generally have less open loop gain at high frequencies so there is less corrective inverse feedback that is available to linearize them. This is especially true of unity-gain compensated operational amplifiers, but is also generally true.

• Woodinville
DSD vs PCM, bit vs hz
##### Reply #52 – 17 February, 2014, 03:03:57 AM
Arnold: Your graph is worthless, because you used (mostly) harmonically related frequencies.

So, if I use a multitone that consists of lines at Fo(2n+1) where Fo is a base frequency, oh, say 200Hz, n is an integer 0 to the result reaching fs/2, I can't see any IMD in that?

You sure about that?

I'm rather sure you're not so sure.
-----
J. D. (jj) Johnston

DSD vs PCM, bit vs hz
##### Reply #53 – 17 February, 2014, 07:17:56 AM
Yeah yeah. Blame the messenger. Please continue disregarding anything I'll say regardless of how ridiculously precise I have to be.

Arnold: Your graph is worthless, because you used (mostly) harmonically related frequencies. That doesn't work for IMD testing.

I find it ironic that complaints about "Blame the messenger" come in a post that also name-calls a well-prepared piece of evidence as "worthless", and provides no reliable evidence to support that exceptional claim.

DSD vs PCM, bit vs hz
##### Reply #54 – 17 February, 2014, 08:38:51 AM
Arnold: Your graph is worthless, because you used (mostly) harmonically related frequencies.

So, if I use a multitone that consists of lines at Fo(2n+1) where Fo is a base frequency, oh, say 200Hz, n is an integer 0 to the result reaching fs/2, I can't see any IMD in that?

You sure about that?

I'm rather sure you're not so sure.

Perhaps our correspondent does not realize that legacy IM tests such as SMPTE and CCIF were designed to be easy to implement with the relatively primitive test equipment of the day (SMPTE 1941, and CCIF/ITU  1929).

In addition the SMPTE test is contrived to work well with a relatively dirty medium - movie film optical sound.

According to http://www.rane.com/note145.html

"Required Conditions. Many variations exist for this test (CCIF). Therefore, the manufacturer needs to clearly spell out the two frequencies used, and their level. The ratio is understood to be 1:1."

IOW the frequencies commonly used for this test (19 KHz & 20 KHz)are not sacrosanct.

• Woodinville
DSD vs PCM, bit vs hz
##### Reply #55 – 18 February, 2014, 06:54:17 AM
"Required Conditions. Many variations exist for this test (CCIF). Therefore, the manufacturer needs to clearly spell out the two frequencies used, and their level. The ratio is understood to be 1:1."

IOW the frequencies commonly used for this test (19 KHz & 20 KHz)are not sacrosanct.

And, of course, with modern FFT analysis one does not, and should not, limit one's self to 2 frequencies.
-----
J. D. (jj) Johnston

DSD vs PCM, bit vs hz
##### Reply #56 – 18 February, 2014, 07:48:56 AM
"Required Conditions. Many variations exist for this test (CCIF). Therefore, the manufacturer needs to clearly spell out the two frequencies used, and their level. The ratio is understood to be 1:1."

IOW the frequencies commonly used for this test (19 KHz & 20 KHz)are not sacrosanct.

And, of course, with modern FFT analysis one does not, and should not, limit one's self to 2 frequencies.

Agreed. As Audio Precision points out a well-designed multitone can affirm excellent frequency response, low nonlinear distortion, and low noise with just one measurement.

http://www.ap.com/kb/show/60

DSD vs PCM, bit vs hz
##### Reply #57 – 19 February, 2014, 04:26:12 AM
The main problem is that you used so many tones, you're obscuring actual intermodulation products. The FFT window used is not selective enough, should use a rapid falloff flat top window, not Hamming.

Also you should perhaps try a less state of the art device. Specifically one that has stated SMPTE IMD at -110 dBV @ 1V RMS, expected CCIF IMD at -90 to -100 dBV @ 1V.
(Of course I'd love to know how they managed to pull it off - must be really careful board layout and power filtering.)

See some sample measurements (the nwavguy's are better):
http://nwavguy.blogspot.com/2012/04/odac-released.html
http://scientistsaudio.blogspot.com/2012/0...kii-part-1.html

That is -106 dBV @ 1V RMS for ODAC, -100 dBV @ 1V RMS for Leckerton from another source I could find. Despite that, it is discernible in a test from Leckerton UHA-6s mkII when used as DAC, same amplifier (FP10). I've done that test way earlier, also single blind, USB source from the same machine. (8/8) So it's likely not IMD. THD dominates IMD in these two devices, but is still very low at -90-something dB.
Could be power supply filtering? But how?

Could be the channel imbalance in that test? But not in this last one...

--
We still need to run listening tests - we do not know *everything* about electronic sound devices yet. Of course, first run the known electrical tests to ensure the devices are not broken.
This includes: jitter (for DACs), IMD, THD, noise, clipping, aliasing and frequency response.

In this test, I have not excluded any of the above except noise and FR, though all are assumed good. I do not have a good enough hardware to measure such low distortion. (Scratch that, I do have it since very recently. Will try to do that next month.)

--
The idea is to explain the difference. Might be the single blind trial process, who knows? I'll have to run double-blind, this is hard with physical devices.
Or maybe I got really lucky and hit 11/12 by chance.
ruxvilti'a

DSD vs PCM, bit vs hz
##### Reply #58 – 19 February, 2014, 08:49:24 AM
The main problem is that you used so many tones, you're obscuring actual intermodulation products. The FFT window used is not selective enough, should use a rapid falloff flat top window, not Hamming.

It appears that some take me for an naive person  that would not check such things out before basing a web site (www.pcavtech.com) or even a post on an important conference such as HA on it. IMO it is similarly unwise act like reliable authorities such as Audio Precision would make the same mistake.

Here is the pure multitone:

Here is the multitone with some nonlinear distortion added:

Obviously, there are more than enough actual intermodulation products to be seen and analyzed.  While the FFT software I am currently using does not support flat topped windowing I am familiar with such things, and their major benefit is that the  tops of the peaks of the multitone spectra would be more even given that the frequency choices were arbitrary. If one chooses frequencies that correspond exactly to FFT bucket center frequencies this minor situation  goes away with any reasonable windowing technique. Of course one has to have some basic understanding of windowing to anticipate this.

• saratoga
DSD vs PCM, bit vs hz
##### Reply #59 – 19 February, 2014, 10:42:04 AM
That FFT argument is pure WTF. The choice of windows determines how precisely you can resolve two different frequencies. If, as we have established, no IMD frequency exists at all the FFT window makes no damn difference.

Its not even wrong, it makes me wonder if you understand what you're even trying to talk about at all.

• pdq
DSD vs PCM, bit vs hz
##### Reply #60 – 19 February, 2014, 11:09:22 AM
That FFT argument is pure WTF.

For those of us who don't understand all of this, whose FFT argument are you referring to, AstralStorm's or Arnie's?

• saratoga
DSD vs PCM, bit vs hz
##### Reply #61 – 19 February, 2014, 11:46:29 AM
That FFT argument is pure WTF.

For those of us who don't understand all of this, whose FFT argument are you referring to, AstralStorm's or Arnie's?

Sorry, I mean AS's argument that the choice of window function would somehow create energy at frequencies where there is none currently. Changing the window might turn one peak into 2, or two into 1, but it still has to conserve energy.  It can't pull it from thin air.

• pdq
DSD vs PCM, bit vs hz
##### Reply #62 – 19 February, 2014, 12:13:49 PM
Sorry, I mean AS's argument that the choice of window function would somehow create energy at frequencies where there is none currently. Changing the window might turn one peak into 2, or two into 1, but it still has to conserve energy.  It can't pull it from thin air.

Thanks - that explanation makes sense to me.

DSD vs PCM, bit vs hz
##### Reply #63 – 19 February, 2014, 02:09:07 PM
That FFT argument is pure WTF. The choice of windows determines how precisely you can resolve two different frequencies. If, as we have established, no IMD frequency exists at all the FFT window makes no damn difference.

Its not even wrong, it makes me wonder if you understand what you're even trying to talk about at all.

Methinks that many people are impressed with NWAVGUY's  fine site, and some without understanding the background think that everybody should do things his way or be in error.

DSD vs PCM, bit vs hz
##### Reply #64 – 19 February, 2014, 04:16:59 PM
It's either that FFT thing or you're facing relatively high jitter in the signal source, therefore the high sidebands after integration.
The whole idea is to have as close to a specific frequency and not have the frequency move around, as the IMD products will also move and reduced by time averaging, biasing the result down.
(Much like spread spectrum is used to reduce noise peaks.) But then, these products, if any, would also get wider.

The window function does not create energy, but the sidebands do obscure any low energy in the signal.
The right window for this, again, is a flat top window with high order sidelobe drop (there do exist 7th order ones) and decent first sidelobe level. There's a tradeoff involved between main lobe width and first sidelobe volume of course.

--
I have another theory: the tube amplifier introduces distortion that interacts with the low level distortion present in the amplifier output, making it more audible and thus discernable.
Plausible?
ruxvilti'a

• saratoga
DSD vs PCM, bit vs hz
##### Reply #65 – 19 February, 2014, 09:35:03 PM
You're mistaken.  800+900 = 1700Hz

From his graph you can see that IMD is less than -129 dB.  Clearly this is negligible.

The whole idea is to have as close to a specific frequency and not have the frequency move around, as the IMD products will also move and reduced by time averaging, biasing the result down.
(Much like spread spectrum is used to reduce noise peaks.) But then, these products, if any, would also get wider.

How many bins would you have to average to get audible IMD down to -129dB?  100,000?  1,000,000?  More?

Seriously, think before you post.

DSD vs PCM, bit vs hz
##### Reply #66 – 20 February, 2014, 02:11:25 AM
How many bins would you have to average to get audible IMD down to -129dB?  100,000?  1,000,000?  More?

Seriously, think before you post.

A bit unstable signal generator clock. (SNR of 1010LT D/A is around 102 dB, so you have many times time averaging already.)
2x averaging drops noise floor by 1-3 dB depending on how white the noise is, drops IMD with a stable clock less. With unstable clock, you can treat IMD as a colored noise.
To go from 102 dB to 128 dB, you need few hundred times averaging, maybe even 1000x. With an unstable signal source that would drop IMD a lot.
(Also, what is with -30 dB up top? Usually -3 dB or -1 dB is used.)
ruxvilti'a

DSD vs PCM, bit vs hz
##### Reply #67 – 20 February, 2014, 09:48:02 AM
How many bins would you have to average to get audible IMD down to -129dB?  100,000?  1,000,000?  More?

Seriously, think before you post.

A bit unstable signal generator clock. (SNR of 1010LT D/A is around 102 dB, so you have many times time averaging already.)
2x averaging drops noise floor by 1-3 dB depending on how white the noise is, drops IMD with a stable clock less. With unstable clock, you can treat IMD as a colored noise.
To go from 102 dB to 128 dB, you need few hundred times averaging, maybe even 1000x. With an unstable signal source that would drop IMD a lot.
(Also, what is with -30 dB up top? Usually -3 dB or -1 dB is used.)

The above seems to paint the picture of a frightful lack of familiarity with FFTs for someone who apparently wants to want to lecture HA about them.

The noise floor for  a random noise as portrayed on a FFT is partially based on bin size.  The SNR that is used on equipment specs is usually based on a 20-20K -3 dB band or A weighting or such. The FFTs I have been posting are based on a 65k point FFT as they self-document.  The bin width is around 1 Hz and depends on the number of points and the sampling frequency.  The amount of noise captured by a measurement is proportional to the square root of the bandwidth of the measurement.  So a measurement taken over a 1 Hz band captures about 140 times less noise or about 43 dB less noise than one taken over a 20-20KHz band. Therefore the noise floor for these charts can be about 43 dB less than the specified SNR of the Delta 1010LT audio interface.

Coherent (steady, pure) signals are different. They typically are what they are as the carrier by definition primarily fits into a single bin.  In general, jitter just makes new sidebands that show up in new bins, it doesn't take much energy out of the carrier.

Knowledge the threshold of hearing and of spreading functions associated with spectral masking strongly suggests that any spurious response that is 100 dB or more down will unconditionally fall under either the masking curve or the curve associated with the threshold of hearing. IOW, it won't be heard, period.

• saratoga
DSD vs PCM, bit vs hz
##### Reply #68 – 20 February, 2014, 04:23:23 PM
How many bins would you have to average to get audible IMD down to -129dB?  100,000?  1,000,000?  More?

Seriously, think before you post.

A bit unstable signal generator clock. (SNR of 1010LT D/A is around 102 dB, so you have many times time averaging already.)

Ok, I see that the problem here is that I was assuming you understood how to read an FFT.  Do you understand Arnold's (very gracious) explanation of what that plot is showing?  If not, you should ask questions until you do.

(Also, what is with -30 dB up top? Usually -3 dB or -1 dB is used.)

Once you understand Arnold's reply, you will be able to answer this question yourself.

(Hint:  the FFT as implemented here conserves power)

DSD vs PCM, bit vs hz
##### Reply #69 – 20 February, 2014, 05:39:05 PM
So yes, it is averaging a lot (because the FFT is long window - I originally thought a shorter window with major additional averaging specifically because of the sidebands - but that wouldn't reduce the level of the tones either, my mistake). 65k FFT is likely around a second of averaging.

65k FFT should be awfully selective in frequency domain, therefore the only reason for sidebands is that the generator clocks are just not stable enough. Therefore any measurement of THD with them will likewise get smeared out and reduced in level... I do hope they're unstable in the same way in all frequencies (say, digital jitter) - otherwise the IMD will indeed average out as it will show multiple different interference frequencies changing around, looking exactly like colored noise, nicely averaging out due to the long window.

I agree about the masking - do note though that this graph pertains to masking with simple signals and did not evaluate timbre on complex signals. Generally there's another effect in play - an interference changes perceived pitch, even if the tone is masked, as long as there is enough of it in total. The ear will "synthesize" a virtual pitch based on the ratio of interferences.
That said, I think the effect is minor as long as the pure tones are audible, especially for an untrained listener. And perhaps even negligible for a trained listener. We'd have to test it, I'm not aware of any research on masking of inharmonic distortion. This specific effect has not been exactly studied, others have.
Also the IMD is typically further away and also below the tones that could mask it - it's similar to higher order harmonic distortion in this way.

Some source on perception of nonlinearities: http://hephaestusaudio.com/media/2008/11/d...rtion_aes_i.pdf
This discusses harmonic distortion mostly...
http://www.gedlee.com/downloads/The%20Perc...0Distortion.pdf <- very good tests.
The distortion is more complex than the numbers would lead us to believe. Again more the reason for listening tests everywhere.

--
The real question still stands. What is the cause for the difference I heard? Is that "tube amp making source distortion more audible due to its own distortion" plausible? Do I get to retry with double-blinding?
ruxvilti'a

• saratoga
DSD vs PCM, bit vs hz
##### Reply #70 – 20 February, 2014, 05:45:54 PM
So yes, it is averaging a lot (because the FFT is long window - I originally thought a shorter window with major additional averaging specifically because of the sidebands - but that wouldn't reduce the level of the tones either, my mistake). 65k FFT is likely around a second of averaging.

I think the problem here is that people were assuming you had some kind of background in engineering/electronics/signals and I see now that you do not.  So what exactly is your background?  We're probably just talking past you with a lot of engineering in this thread.

The real question still stands. What is the cause for the difference I heard? Is that "tube amp making source distortion more audible due to its own distortion" plausible? Do I get to retry with double-blinding?

bad testing, bad equipment, bad luck, etc

This is the fun part of hardware testing.  Something gives you an unexplained result.  Now you get to dig into it and try to figure out what it is.

DSD vs PCM, bit vs hz
##### Reply #71 – 20 February, 2014, 05:53:36 PM
Pah, I just need to read the actual images closely. It has the bin size and windowing function spelled out...
Huge wall of texts would've been avoided if I actually *looked* at the images closely.

Blame me for replying in the morning/late in the evening. Now I feel stupid.

So the only remaining thing is to use a more stable signal source and null out the sines for readability.

--
Yeah, I will go digging for the cause. Starting next months with more advanced measurements of all parts of the chain. Either I'll have to lift the new Agilent - likely impossible, someone will notice and stab me - or carry expensive equpiment on my back to that lab and hope it's nothing power supply related.
ruxvilti'a

• saratoga
DSD vs PCM, bit vs hz
##### Reply #72 – 20 February, 2014, 06:05:13 PM
Pah, I just need to read the actual images closely. It has the bin size and windowing function spelled out...

Well obviously that will help, but its not the only problem here.  You need to understand the images too.  All the data in the world won't help you if you don't understand what it means...

Yeah, I will go digging for the cause. Starting next months with more advanced measurements of all parts of the chain. Either I'll have to lift the new Agilent - likely impossible, someone will notice and stab me - or carry expensive equpiment on my back to that lab and hope it's nothing power supply related.

I would start much more simple than that.  A very basic RMAA test on an ordinary PC will give you a good idea if something is wrong with your hardware.  No need to use complex equipment yet.

• greynol
• Global Moderator
DSD vs PCM, bit vs hz
##### Reply #73 – 20 February, 2014, 06:26:08 PM
So what is all the fuss about here, different hardware measuring differently and a questionable listening test that confirmed it?

Rather than dive into obscure faux-technical excuses (jitter is the ultimate boogeyman), perhaps we can focus on the latter.  Samples demonstrating audibility would be in order as well.
Please be aware that there is an expectation that you read and follow the rules.  It is not my job to teach them to you.

Your eyes cannot hear.

DSD vs PCM, bit vs hz
##### Reply #74 – 21 February, 2014, 03:23:55 AM
I would start much more simple than that.  A very basic RMAA test on an ordinary PC will give you a good idea if something is wrong with your hardware.  No need to use complex equipment yet.

No, it won't. RMAA is quite incomplete in many ways. I'd also need a soundcard with input that is as good or better than DUT.
I'm not sure I can trust the Lynx for this yet, since I haven't measured its input. (non-loopback) FP10 has known broken inputs, nonlinear in phase and FR.

(I can make RMAA handle absolute levels by calibrating the level beforehand with the voltmeter.)

http://nwavguy.blogspot.com/2011/02/rightm...lyzer-rmaa.html - some concerns.
A very long time ago I've ran a test with 3 different sets of volume levels on an Aureon 7.1 Space. The results were wildly different, esp. crosstalk.

I'll mention the more known examples of what I listened to during the test:
Faster than Light Soundtrack - Engi (battle); background aharmonic tone - sounds more/less atonal
PPPPPP - Pressure Cooker; noise timbre
Kraftwerk - Tour de France - Titanium; metallic attack timbre, esp. in presence of thicker bass near the end

The differences are subtle. I can isolate some of the problem spots.
Of note, these tracks also overcome LAME even at a.p.e. Vorbis manages at q6+.

As I've mentioned, I used Hifiman HE-500 w/ pad and grille mods, equalized to perceptually flat with tones and modified double-precision Calf parametric equalizer - the headphone is quite low distortion. Not sure yet how much distortion the tube amp (Eddie Current Super 7) has, but supposedly low as well...

--
Oh, there's a factor that could be vastly different - crosstalk...
ruxvilti'a