Alternative method for measuring distortion

[j_j]

At this stage/level of the discussion such “general” arguments are not enough. Please, elaborate - what exact premise you mean, why is it incorrect and provide the correct one.

You refuse to accept the idea of "information" and continually reject that as a meaningful issue. This is really tiresome.

Once again, it is malpractice and false teaching to claim that the purpose of dither is to help perception.

Dither retains more information than the lack thereof. It's that simple. Live with it.

As to this:

Exactly, adding randomness decreases information, not increases it.

It was shown, irrevocably and without any possible dispute whatsoever in any way, shape, or form, to be utterly, completely, and without any possible revocation, to be wrong, in the thread on dither.

Retract your false teaching and STOP IT. You sully the field with your misinformation.

As to your 'randomness' canard, which has more information content, a signal that has a flat spectrum, or a signal that has a heavily shaped spectrum?


Do tell, @Serge Smirnoff please.

Let us, take for instance, the utterly irrevocable disappearance of a sine wave in an undithered state when it sinks below +-.5 in a zero-centered quantizer. Let us also encode our information in the level of that sine wave. When the sine disappears, the information is gone forever.

In the case where somebody DOES IT RIGHT, ACCEPTS THE MATH, and adds dither before quantization, now the level is easily detected in a long FFT (for instance, there are better ways, but this isn't a signal detection class) and the information is preserved to infinitely more resolution.

(to explain, zero information without dither, so the information remaining, while impaired compared to the original, is a ratio of information preserved of finite number divided by zero, or infinitely better signal preservation.)

ENOUGH!!!!

 

[j_j]

For sure, linearization is mathematical, but its use in audio is determined by psychoacoustics.

The fact that it also addresses psychoacoustics does not make it any less a required use FOR LINEARIZATION. See my previous post wherein information retention proves your mistake.

 

[j_j]

For sure, and is not required in another ones.

Places dither is not required:
1) When other noise is known to be substantially larger. (slightly increased or decreased dither can actually harm the effect of dither)
2) When attempting to reject information, as in an MP3 encoder, etc.
3) Other places where reducing information is desired. There are a few specialized places.

 

[xr100]

Places dither is not required:
1) When other noise is known to be substantially larger. (slightly increased or decreased dither can actually harm the effect of dither)

Could you elaborate, please?

 

[j_j]

Could you elaborate, please?

For levels of dither near 1 LSB, you want exactly TPD at 2 LSB peak.

To explain, it's easier just to generate TPD of slightly different values and then look at the noise in a very slowly varying signal. But once you get a few times the LSB larger, you're cool. Notice, if you will, how the dither effect comes and goes (not completely, but to some extent) as the levels move near the center of the quantizer value. It's a small effect, but this is why at a factor of close to 2 it's mostly gone, and at 4 it's gone. BUT the point holds, there is an "exactly right' quantity of dither for minimum SNR impairment and constant noise density across time. Oh, this is 1.1 of the proper dither.

 

[abm0]

Measurement conditions can be discussed. Output level I used is max. level allowed for DAPs by EU standard (hearing loss ...), load now is resistive but can be complex (real headphones with Olive target response or, even better, an electric equivalent of such phones).

OK, can be, but has not been for results published so far. So would you agree that existing published Df-values for your list of devices may be inconclusive for everyone not planning to use them to power transducers almost identical to the loads in your test setup? For example even the Hiby FC3 which is the subject of some hype on your website - and on HypeTheSonics - I have found to produce impactless bass and "body-less" voices already through a 35-ohm HE-400i, which is electrically very similar to your 32-ohm load (but has not-great efficiency). Maybe there should've been more information attached to the metric report about how quickly those results could diverge with changing load characteristics and less hype about the device supposedly being already an example of perfect engineering for the whole market to use as a reference?

 

[IAtaman]

Exactly, adding randomness decreases information, not increases it.

Sounds like it doesn't it. Important to remember though you are adding controlled randomness before truncation, which improves the linearity of the truncation process itself. I would suggest to read about dither from a non-audio source. Many resources who try and explain what dither is from audio perspective do not do very a good job to be honest, and some of them are down right wrong, I don't know why. Maybe dither is hard to explain with an audio signal as example? Anyway, personally I found sources that use image processing examples do a much better job explaining what is dither and how it helps to preserve information.

 

[solderdude]

Talk about thread necroing .... this thread kind of 'stopped' more than 3 years ago....

 

[IAtaman]

Talk about thread necroing .... this thread kind of 'stopped' more than 3 years ago....

Ha! Did not realize that. The arguement was so heated if felt like its just out of the oven. I blame @abm0.

 

[abm0]

Well I'm still interested in the Df-metric itself and the cheap devices it made to appear as being giant killers, i.e. the Hiby FC3 or Shanling M0 or CX31993, and what major deficiencies of the method we can find that would invalidate those conclusions and the associated hype. (No, I didn't necro this for the long off-topic about dither, which already had a better home on another thread.)

 

[solderdude]

Give the PK metric a try.
Make sure the used sound-card is top notch otherwise it will impart its signature as well.

 

[abm0]

Oh I'm not looking to redo the measurements for myself, I want to clarify the theoretical/methodological limitations behind the already published results. For example if the FC3 has "Sabre glare" yet the Df-metric shows it to have the absolute best aggregate result of all devices tested so far (including Chord desktop units), this might highlight - among other issues - one big vulnerability of these single-metric methods, even if the test includes a variety of musical signals as inputs. Since the deficiency only manifests in a specific and narrow part of the spectrum, and some arbitrary portion of the input music collection may not feature that band too prominently, you could end up with an error contribution that's not moving the end result by much, yet it would make many buyers recoil in horror at the thought of spending money to listen to such a device. Not the kind of problem I'd expect to get rid of by simply choosing someone else's singular aggregate metric like the PK, either.

 

[SIY]

For example if the FC3 has "Sabre glare"...

...demonstrated to be real by exactly no-one.

 

[abm0]

...demonstrated to be real by exactly no-one.

Huh? I could've sworn Amir has shown it on several devices' IMD plots? (Of course that still always left me with the question "why doesn't it also show on a pure and simple FR graph for the same device, if it's such an audible problem". As yet unresolved.)

 

[IAtaman]

Huh? I could've sworn Amir has shown it on several devices' IMD plots? (Of course that still always left me with the question "why doesn't it also show on a pure and simple FR graph for the same device, if it's such an audible problem". As yet unresolved.)

Maybe you are thinking of ESS IMD hump? Because in another forum in which you seem to be participating as well, Amir calls Sabre glare folklore.

ESS DACs "Sabre glare" measurements?

Not sure if this is more about a type of source device or more about science and measurements, but here it is: Does anyone know of any measurements showing the existence of the oft-mentioned "Sabre glare" in ESS Technology's Sabre-branded DACs? If there were audible extra content in the high...

www.head-fi.org

 

[abm0]

Maybe you are thinking of ESS IMD hump?

Well, yes. I thought the IMD hump was where the real basis was found for the rumors/folklore bearing the label "Sabre glare", in the years since that old (and very short) Head-Fi discussion.

 

[j_j]

Sigh. There are a bunch of plots in this thread, and another thread written for the actual subject of dither, all of which show conclusively that dither works, and no, it does not decrease "information".

Many audio people do not, to this day, understand the somewhat counterintuitive effect.

 

[SIY]

Huh? I could've sworn Amir has shown it on several devices' IMD plots? (Of course that still always left me with the question "why doesn't it also show on a pure and simple FR graph for the same device, if it's such an audible problem". As yet unresolved.)

Even for the so-called ESS Hump, measurable =/= audible.

 

[j_j]

Even for the so-called ESS Hump, measurable =/= audible.

Well, I have to say, given a standard Audient 96/"24" (yes, that's quotes of contempt there, nothing is really 24, in or out, more like 19 to 21) convertor and a PC, it's easy to measure things that nobody could EVER hear, so the fact one can measure something does not, most assuredly, mean something is audible.

 

[Curvature]

Well, I have to say, given a standard Audient 96/"24" (yes, that's quotes of contempt there, nothing is really 24, in or out, more like 19 to 21) convertor and a PC, it's easy to measure things that nobody could EVER hear, so the fact one can measure something does not, most assuredly, mean something is audible.

I'm a little sad that I'll likely never know enough about audio to evaluate your statement in worthwhile detail.