New Xiph.Org video: "Digital Show & Tell"

[rbbert]

Although the quote from MFSL that Bruce posted used a sine wave as its example

 

[xiphmont]

I think one of the reasons people think that digital creates stair steps is because in the recording stage, it does as you have shown. However, when it goes back to an analog signal after it has been converted by the analog output stage, it's a continuous sine wave. The fact remains it is a stair step when recorded.

No. I was quite clear on this in the video that this is incorrect, and went into detail as to why.

I also think that showing a 1 kHz sine wave after it has been recorded might just be misleading because music is never just a 1 kHz sine wave.

It is not in any way misleading. I'm using a sine wave only because it's easy to see. I also use square waves (again, easy to see).

I think it would be better to show screen shots of full frequency music being recorded digitally and show what it looked like on the scope prior to D/A conversion and what it looked like on another scope after it came out of the D/A converter
to show how closely it resembled the input signal.

I agree actually, but analog-only equipment is not capable of this; you need to go digital for that, and many too many people do not trust digital. So... tackling one thing at a time. For the record, I also show square waves and show they also obey theory perfectly. The square wave in question had nine harmonics that fit under 20kHz.

I don't think that showing a pure 1 kHz sine wave going through the D/A conversion is telling the whole story.

Could you perhaps state a specific objection? The video leaves out a great deal of nuance-- that's unavoidable. It's covering most of an undergraduate semester in 20 minutes. But it is not fudging any truths. Linear superposition is real. Digital sampling behaves as presented regardless the complexity of the waveform.

 

[RBFC]

Great post, Mark. Better measurements may eventually help to increase our understanding of why things sound different/better/worse.

Lee

 

[opus111]

Great presentation. Only quibble is just after 3mins - cheapo consumer grade boxes 'already at the edge of ideal transparency' ? For sine waves, sure but not for music.

 

[xiphmont]

Great presentation. Only quibble is just after 3mins - cheapo consumer grade boxes 'already at the edge of ideal transparency' ? For sine waves, sure but not for music.

Some were, some weren't. In the cases where manufacturers did a bad job on the analog stages or support circuitry, sure, you're right. The ADCs/DACs themselves though?...
You're welcome to arrange a testing/thrashing of the box I used in the vid if you like. It currently goes for $15-30 on eBay (granted it was $300 in 2000.) Note that it has some serious flaws [for bench test use anyway] that are avoidable. In my own testing I found the chip manufacturer spec sheets to match the input/output performance dead on, so you could comment from those if you like.

 

[opus111]

Definitely the DAC chips (in my experience) - for the reasons you hint at around 16mins in. Inadequate dither is being applied to truncation to low bit. Nowadays DACs are almost all S-D type with 5 or 6bit DACs requiring truncation. The result - audible noise modulation. Doesn't show up on sinewaves on scopes, just as your 8bit truncation test didn't show visible scope noise (to me on the vid anyway).

Noise mod does show up on some datasheets (AKM is one) - compare the noise at full-scale output with that at -60dB on FFT plots.

 

[xiphmont]

BTW, pulled up the sheets, the emagic 2|6 is using an AK4527VQ, the spec sheets for the 4527B also match what I measured myself from the eMagic fairly closely in case you can't find plots for the original part. They improved distortion a smidge.

Definitely the DAC chips (in my experience) - for the reasons you hint at around 16mins in. Inadequate dither is being applied to truncation to low bit. Nowadays DACs are almost all S-D type with 5 or 6bit DACs requiring truncation. The result - audible noise modulation.

FTR, HA did a comparison of dithered and undithered truncation a few years back [which I've since been unable to locate-- I've asked them to help me find it] and found that a minority of listeners preferred the undithered truncation, either because or despite the test included fade-outs. So, I'm not sure that alone disqualifies a DAC from being 'musical'. Dither is, IMH[ButDebatable]O, overrated.

[edit: oh, you meant as an implementation detail within the DAC?]

Doesn't show up on sinewaves on scopes, just as your 8bit truncation test didn't show visible scope noise (to me on the vid anyway).

it was noticeable in person, you can just see it in the 720p version [the compression filters some of it out unfortunately]. FTR, I was not running the DAC in an 8 bit mode there (it doesn't have one), I was applying a full-power TDPF dither in software before truncation.

Noise mod does show up on some datasheets (AKM is one) - compare the noise at full-scale output with that at -60dB on FFT plots.

I'll remember that.

 

[opus111]

BTW, pulled up the sheets, the emagic 2|6 is using an AK4527VQ, the spec sheets for the 4527B also match what I measured myself from the eMagic fairly closely in case you can't find plots for the original part. They improved distortion a smidge.

Eyeballing the sheet now - in order to make an estimate for the noise modulation, I have to make an assumption. There are no FFT plots so the assumption can't be checked - the assumption is that the -0.5dBFS THD+N is dominated by noise and not distortion. I've never seen one that isn't but its not 100% secure as a hypothesis.

So we have the DAC's typical THD+N at 90dB close to fullscale, whereas at -60dB we get -106dB. This one's A-weighted so take off a couple of dBs and tentatively conclude that the noise modulation is 14dB. This is an average noise modulation, as your video showed, the noise is going up and down with the signal level. By a similar method we get the ADC's noise modulation at 8dB.

<edit> Yes I meant there's no way that I know of to ensure the correct noise pdf in a noise shaping feedback loop - its intrinsic to the way that the DAC chip's operating, and its a truncation to 5 or 6bits, not 16.

 

[xiphmont]

Eyeballing the sheet now - in order to make an estimate for the noise modulation, I have to make an assumption. There are no FFT plots so the assumption can't be checked.

First hit on Google for "AK4527" has a sheet with plots.

 

[opus111]

Ah thanks - I only got the chip DS from my Googling, not the eval board. Looks like AKM's full-scale THD (without the noise) is around -106dB so the assumption holds.

 

[jmvalin]

I think one of the reasons people think that digital creates stair steps is because in the recording stage, it does as you have shown. However, when it goes back to an analog signal after it has been converted by the analog output stage, it's a continuous sine wave. The fact remains it is a stair step when recorded. I also think that showing a 1 kHz sine wave after it has been recorded might just be misleading because music is never just a 1 kHz sine wave. I think it would be better to show screen shots of full frequency music being recorded digitally and show what it looked like on the scope prior to D/A conversion and what it looked like on another scope after it came out of the D/A converter to show how closely it resembled the input signal. I don't think that showing a pure 1 kHz sine wave going through the D/A conversion is telling the whole story.

Have you actually watched the video? The only stairsteps in there is when Monty arbitrarily decides to draw stairsteps to explain why it's wrong. Both the ADC and the DAC contain filters that would prevent you from having stairsteps (at least any stairstep with steps the size the sampling interval) even it you tried. When bandlimited, the signals aren't a summation of steps, they're actually a summation of very smooth sinc (sin(x)/x) functions.

 

[xiphmont]

So we have the DAC's typical THD+N at 90dB close to fullscale, whereas at -60dB we get -106dB. This one's A-weighted so take off a couple of dBs and tentatively conclude that the noise modulation is 14dB. This is an average noise modulation, as your video showed, the noise is going up and down with the signal level. By a similar method we get the ADC's noise modulation at 8dB.

Except that the majority of that difference is the distortion products disappearing below the noise floor. That's not disagreeing with what you said, but it's not anywhere like the same thing as a 14dB/8dB white noise modulation. it's possibly the most innocuous thing that could happen (short of there being no distortion at all).

 

[opus111]

Perhaps I didn't explain my point clearly enough - I'm here suggesting that an average noise floor modulation figure merely indicates there's a problem in that they're not dithering correctly, its not really saying very much about whether its subjectively audible. The subjective effects I hear (not with this particular part, most recently with a CS4398) are almost certainly above the level of -90dB which I doubt would make any subjective difference.

 

[xiphmont]

Perhaps I didn't explain my point clearly enough - I'm here suggesting that an average noise floor modulation figure merely indicates there's a problem in that they're not dithering correctly

Why would that be the case over some other form of nonlinearity? I would expect faulty dither to result in a distortion amplitude figure that's highly nonlinear with input amplitude. On this part at least, the distortion products fall off smoothly with input amplitude, much like most opamps. I realize you can't see that from the spec sheet, and you'd have to take my word for it.

its not really saying very much about whether its subjectively audible. The subjective effects I hear (not with this particular part, most recently with a CS4398) are almost certainly above the level of -90dB which I doubt would make any subjective difference.

Ah, OK. You'd objected to my statement of old/consumer-grade DACs being suitable for anything but sinewaves, so I assumed you were then offering an example of exactly that (when this example that was about as innocuous/inaudible as they get

 

[opus111]

Why would that be the case over some other form of nonlinearity?

Ooops, you lost me. Noise modulation is known for being a byproduct of the wrong pdf of the dither. Rectangular pdf gives it for example.

I would expect faulty dither to result in a distortion amplitude figure that's nonlinear with input amplitude. On this part at least, the distortion products fall off smoothly with input amplitude, much like most opamps. I realize you can't see that from the spec sheet, and you'd have to take my word for it.

How do you know they do? And what form of distortion products - THD or IMD (just for starters) ? As for opamps, the distortion products go up substantially with frequency (like 18dB/8ve) and its in practice almost impossible to keep the HF out of one's opamp (I know, I've tried).

OK. You'd objected to my statement of old/consumer-grade DACs being suitable for anything but sinewaves, so I assumed you were then offering an example of exactly that (this example that was about as innocuous/inaudible as they get

No, I haven't objected, I've quibbled with your claim of 'near ideal transparency' - because I've experienced most DACs as far from transparent in regards to noise modulation (which I'd gladly trade for some fairly innocuous THD if that's low order coz my transducers are adding that anyway). In order to provide a hypothesis in support of my observations I've pointed out the smoking gun in the DS of the part you were claiming for.

 

[xiphmont]

Ooops, you lost me. Noise modulation is known for being a byproduct of the wrong pdf of the dither. Rectangular pdf gives it for example.

Now I'm confused too. You said that the distortion products of this specific part were likely the result of botched dither. I said that in that case, I would expect a different behavior than the products smoothly falling below the noise floor as the input amplitude was reduced, specifically, that I'd expect a slower falloff until things got to around 1-2LSB and then a large change. I'll admit I've not played much with RDPF dither, so I'm not as familiar with its behavior. Is that what you're suggesting is happening here? If so... then I understand. If not, elucidate.

[I suppose I don't need to ask. I can go downstairs and just measure it myself.]

How do you know they do? And what form of distortion products - THD or IMD (just for starters)

From measuring individual distortion harmonics sweeping both frequency and amplitude.

As for opamps, the distortion products go up substantially with frequency (like 18dB/8ve) and its in practice almost impossible to keep the HF out of one's opamp (I know, I've tried).

I was referring only to the distortion products dropping off with reduction in amplitude much as happens with an opamp. It was a similie, not a suggestion that, eg, the output buffer was responsible.

No, I haven't objected, I've quibbled with your claim of 'near ideal transparency'
[...]
In order to provide a hypothesis in support of my observations I've pointed out the smoking gun in the DS of the part you were claiming for.

What about this part does not meet the criteria of 'near ideal transparency'? [and yes, I realize your quibble was only that-- a quibble. I forgot to thank you for an otherwise positive review :-]

 

[opus111]

Now I'm confused too. You said that the distortion products of this specific part were likely the result of botched dither.

No, unless I'm being cryptic again without realizing it (always a possibility). I only pointed to the FFT containing the distortion products (which you kindly found for me) as a way to verify that the full-scale THD+N measurement was noise dominated. So I was pointing to the relative absence of distortion products and the presence of (undesirable) noise products.

I'll admit I've not played much with RDPF dither, so I'm not as familiar with its behavior. Is that what you're suggesting is happening here? If so... then I understand. If not, elucidate.

I'm suggesting that the noise modulation being generated (not in dispute is it?) is the result of incorrectly dithered truncation occurring at the 5 or 6 bit level. The DS doesn't say how many bits in the DAC part so I'm guessing here. Whereas I probably agree that dither at the 16bit level might well be overrated (for me the jury's still out) at the 6bit level I'm in no doubt, its audible.

What about this part does not meet the criteria of 'near ideal transparency'?

I'm surmising its the incorrectly dithered truncation at the 5 (or 6) bit level. But as I've not heard the part in question, its just conjecture.

 

[xiphmont]

No, unless I'm being cryptic again without realizing it (always a possibility). I only pointed to the FFT containing the distortion products (which you kindly found for me) as a way to verify that the full-scale THD+N measurement was noise dominated. So I was pointing to the relative absence of distortion products and the presence of (undesirable) noise products.

Ah. I'd seen [and had pointed out] relatively little modulation of the non-distortion noise products (~3 or 4dB between 0dBFS and -60dBFS in the DAC, and then no additional change below that), so I hadn't considered that as something you were talking about.

I'm suggesting that the noise modulation being generated (not in dispute is it?) is the result of incorrectly dithered truncation occurring at the 5 or 6 bit level. The DS doesn't say how many bits in the DAC part so I'm guessing here. Whereas I probably agree that dither at the 16bit level might well be overrated (for me the jury's still out) at the 6bit level I'm in no doubt, its audible.

OK, I understand. I would not personally think a 4dB noise floor modulation around FS that's affecting a critical band's noise floor 120+dB down (100dB down broadband) to have any chance of being audible, period. But I'll admit-- I've not run that specific test. If it _is_ audible, I still find it hard to object to.

I'm surmising its the incorrectly dithered truncation at the 5 (or 6) bit level. But as I've not heard the part in question, its just conjecture.

I can send you one if you like. I have a pile of spares for when I accidentally blow one up in the lab. I've lost a couple that way. But really, I don't think you'd learn anything not in the sheet. By having one I mean, not by blowing it up.

 

[opus111]

OK, I understand. I would not personally think a 4dB noise floor modulation around FS that's affecting a critical band's noise floor 120+dB down (100dB down broadband) to have any chance of being audible, period.

Sure, can't disagree - but we only have the average noise floor (sine stimulus) reading gleaned (by deduction) from the DS, not the instantaneous noise floor modulation with high crest factor signal. Which I'm betting is what I'm hearing and dlsliking and hence it must be considerably higher than -100dB down on music I reckon.

I can send you one if you like. I have a pile of spares for when I accidentally blow one up in the lab. I've lost a couple that way. But really, I don't think you'd learn anything not in the sheet. By having one I mean, not by blowing it up.

Its a kind offer but I'll pass as I'm already overloaded with sucky-sounding D-S (or S-D) DACs

 

[xiphmont]

Sure, can't disagree - but we only have the average noise floor (sine stimulus) reading gleaned (by deduction) from the DS, not the instantaneous noise floor modulation with high crest factor signal. Which I'm betting is what I'm hearing and dlsliking and hence it must be considerably higher than -100dB down on music I reckon.

That should be a relatively easy test to run since all you're looking for is the noise figure. Got any specific examples of music in mind?

Its a kind offer but I'll pass as I'm already overloaded with sucky-sounding D-S (or S-D) DACs

Buuuurrrrrn. Sounds like you want out ;-) OK, you can have the parting shot.