Lavry Engineering Paper on Hi-Res

[Jud]

...may want to have a long talk with Dan Lavry. His DAC (the DA11) will accept S/PDIF or AES input at sample rates up to 200kHz. (USB and optical are limited to 96k.) The DAC chip he uses employs the standard "8x oversampling" ( http://www.analog.com/static/imported-files/data_sheets/AD1955.pdf ) so in fact the sample rate at the point where the D/A conversion is done is 384 or 352.8kHz, depending on input sample rate.

 

The only difference the input sample rate *may* make in Dan Lavry's DAC is that a higher rate *may* avoid one or two 2x oversampling steps. (I'm not conversant enough with electronics to know from the DAC chip info sheet. Maybe someone else can tell us if interested.)

 

[Jud]

But, what I get out of the "general message" from his thesis is 196 and above might be out of range of human hearing.

 

44.1 is out of the range of human hearing, too. The frequency response issue is just one aspect of the overall sound quality question. Some designers and peer-reviewed papers think 192kHz makes a difference re frequency response, others don't. But most of the designers who talk about hi-res don't tend to mention frequency response as a major consideration. Rather, they talk about filtering.

 

Now Lavry says in his paper that once you're up to about 60kHz, you've got no more worries re filtering. But here's what Keith O. Johnson of Spectral says in an interview in The Absolute Sound ( http://www.spectralaudio.com/TAS_%20Interview.pdf ):

 

How has having access to the 176kHz/24-bit high resolution files of your own recordings affected your design work on playback equipment?

 

KJ It’s affected it very, very much. The high-res files set a gold standard against which you can judge CD playback. I should mention that CD can be very good, indeed. We think, “Oh my gosh, it’s 44.1kHz and 16 bits—how can that be even in the same league as something that’s got ten times more information?” It turns out if one works a CD very, very well, with the proper noise-shaping system that is accurate to about 1 part per million, then the remaining difference between high resolution and the 44.1 is time dispersion. The other factors really are not that important. And the time dispersion can be partially corrected by doing group-delay corrections.

 

Time dispersion, for the benefit of the readers, is a

smearing of transient information that occurs in digital

filters. High sampling rates relax the filter requirements

so there’s less of this spreading out of transient energy

over time.

 

KJ Exactly. That’s exactly what happens, and it’s very technical, but it’s highly audible.

 

* * *

 

So who to believe, Keith O. Johnson or Lavry? Well, besides the fact that KOJ designs the best audio electronics I've ever heard (Spectral Audio) and has won Grammys for audio engineering, Lavry's paper leaves me cold on a couple of counts:

 

1 - There's some "sleight-of-hand" going on. You're right, his general message appears to be about frequency response, but no one is really trying to argue Shannon-Nyquist. The more important issue is filtering, and on that Lavry seems more cagey than open.

 

2 - About filtering: Lavry talks about 60kHz being good enough, and later "proves" his point by illustrating a couple of filters (Butterworth and Chebyshev) with "no ringing." Except there is no such thing. The sharper a filter is, the more pre-ringing and post-ringing it has. ( http://en.wikipedia.org/wiki/Gibbs_phenomenon#Signal_processing_explanation ) This is a proven mathematical and physical reality, just like Shannon-Nyquist. Higher frequencies have the advantage of allowing gentler filtering, but since Lavry is set on saying 96kHz is good enough, he slides right past that point.

 

3 - Lavry's major objections to encoding at resolutions above 96kHz appear to be these: (a) unduly large files; (b) unduly high CPU resources; © unduly high equipment expense; (d) lower dynamic range at higher resolutions, which he presents as a "permanent" (immutable) law.

 

We can dispose of a, b and c pretty quickly. Download speeds and CPU capabilities have increased, and prices have decreased, to the point that 24/192 ADC capabilities are available in cheap consumer sound cards and downloads of high-res files take a few minutes. But wait - 24/192 consumer-level ADC? I thought Lavry said it was an immutable law that higher resolutions meant lower bit rates.

 

Look more closely, and once again Lavry appears to be playing fast and loose with the facts to try to make his point. He says at 100MHz sample rates only 8 bits of dynamic range are possible. So who exactly is holding out for a 100mHz sample rate? He then says 16 bits are possible at 1mHz. Again, who is asking for such a dynamic range/sample rate combination? (SACD/DSD is recorded at 2.8/5.6mHz with "1-bit" dynamic range. Do you notice the designers here saying this is inadequate for decent sound?)

 

Surely he will discuss the sampling frequencies of interest, 176.4kHz and 192kHz, right? Errrmm, no. He drops right down to 50-60Hz as his next example. So in a 27 page paper supposedly devoted to showing why 96kHz is better than 176.4 or 192kHz, he never actually points to anything objectionable about the latter two sampling rates. (Clearly 24 bit/192kHz parts are possible - as I noted above, one can find them in consumer sound cards for quite reasonable prices these days, to say nothing of professional recording equipment.)

 

So thanks, but as between KOJ and Dan Lavry, I'll take the guy who *doesn't* seem to be trying to obfuscate.

 

[Jud]

Spectral, as good as it's reputation and quality is, leaves me cold, un-emotional, if you will.

 

That's cool. There are many folks with the budget to choose between Spectral and other components who choose something else.

 

But KOJ has done a hell of a lot of research into digital recording and playback, and his discussion of filtering and transient response reflects what I've read elsewhere. Mani mentioned the lack of this in Lavry's paper, and I went into more detail about it, including quoting KOJ re why it is important.

 

[Jud]

If you've already answered this elsewhere, I can go look for it: What would you say are the main ways in which 2x rates through good equipment fall short of your mic feed?

 

[Jud]

One of [Lavry's] basic points, near the beginning, is that you don't get anywhere near a 24-bit word length due to inherent inaccuracies until you have a sample rate as low as 50-60 Hz.

 

But several people here are totally ignoring this and talking about 24/192.

 

So do you think he is just plain wrong on this?

 

Let's do the furthest thing from ignoring, and see where paying close attention gets us.

 

Here's what the paper says about sampling rates and word length:

 

On page 1 -

 

There is also a tradeoff between speed and accuracy. Conversion at 100MHz yields around 8 bits, conversion at 1MHz may yield near 16 bits and as we approach 50-60Hz we get near 24 bits.

 

And on page 27 -

 

AD converter designers can not generate 20 bits at MHz speeds, yet they often utilize a circuit yielding a few bits at MHz speeds as a step towards making many bits at lower speeds.

 

Did you see anything there comparing bit depth at 88.2 or 96kHz to bit depth at 176.4 or 192kHz, or referring to 192kHz at all, 192kHz being the sampling rate Lavry announces on page 1 he will prove wanting ("...the author's motivation is to help dispel the wide spread misconceptions regarding sampling of audio at a rate of 192KHz")? He's got data points at 100mHz, 1mHz, and 50-60Hz. Sorry, but I find his three data points over a 100mHz range - his "sampling rate," if you will - lacking in terms of telling us what the max bit depth is at 192, or comparing this to 88.2 or 96kHz.

 

There are two other points I'd like to make here:

 

1 - Does Lavry's "permanent" bit depth limit actually exist?

 

Lavry says "The compromise between speed and accuracy is a permanent engineering and scientific reality." He has quantified this "compromise" - the bit depth limits - twice in his paper. So presumably we are to take as permanent scientific reality the reference to not being able to get to 24 bits until we "approach" 50-60Hz (though how close that approach must be is a mystery, since Lavry's next data point is at 1mHz).

 

But Barry Diament talks about using bit depths of 64 and 80 while working on his 192kHz sample rate recordings. And you or I can buy inexpensive consumer sound cards that will encode audio at 24/192. (I did, purchasing a very well thought of card for under $200 - http://www.esi-audio.com/products/julia/ . You've previously referred to UK laws about false advertising, so I note in that connection the card is for sale in the UK.) So perhaps Lavry's specific bit depth limitation is not such a permanent feature of the scientific and engineering landscape after all.

 

2 - Does Lavry fairly characterize a limitation in bit depth as a limitation on the "accuracy" of the audible signal?

 

Eight times in the 27-page paper, Lavry characterizes bit depth limitation as reducing the "accuracy" of the musical signal that can be obtained from a recording with a 192kHz sampling rate, though he never says what the bit depth is at 192kHz, nor does he compare it to 88.2 or 96kHz. Lavry laments the inaccuracy resulting from only 16 bits at 1mHz.

 

16 bits at 1mHz is a bit rate almost three to six times higher than SACD/DSD. 8 bits at 100mHz is almost 150 to 300 times higher. Lavry's bit depth limitation, if it existed and if it in fact reduced the accuracy of the musical signal, would quite simply render reasonable-sounding SACD/DSD recordings impossible. Since SACDs and DSD files seem to exist; many people who have an interest in good audio seem to like them; and knowledgeable programmers/designers don't seem to have any conceptual problem with how SACD/DSD recordings work; then my conclusion is Lavry's eight references to problems with "accuracy" are hogwash.

 

[Jud]

I very much appreciate Lavry Engineering's explanation of what was intended by the paper regarding the topic of bit depth.

 

So the assertion that I said there is a "permanent bit depth limit” is not true.

 

I apologize if there is anything in the article about which I gave a misimpression. That wasn't my intention. However, please note that both Mark Powell (in a comment favorable to what he took to be a point made by the article) and I were left with the idea that the bit depth quantifications given in the article were supposed to be valid today. Mark's comment was:

 

One of [Lavry's] basic points, near the beginning, is that you don't get anywhere near a 24-bit word length due to inherent inaccuracies until you have a sample rate as low as 50-60 Hz.

 

But several people here are totally ignoring this and talking about 24/192.

 

So do you think he is just plain wrong on this?

 

There is nothing in the article saying "Currently...," "At the time of writing...," etc., to denote that the quantifications regarding bit depth were only intended to apply as of 2004. The only statement in the article regarding duration is the one I quoted, "The compromise between speed and accuracy is a permanent engineering and scientific reality." Perhaps a more apt phrasing for what was intended would be "While the bit depth attainable at a particular sample rate has increased over time and may be expected to continue to do so, a given technology will always be capable of greater bit depth at lower sample rates."

 

If anyone at Lavry Engineering would like to respond regarding the topic of bit depth in a historical context, I would be interested in knowing what bit depths were commonly used in recorded material available at the "2x" sample rates (88.2kHz and 96kHz) and the "4x" ones (176.4 and 192kHz) when the paper was written, and what the corresponding bit depths are today. I would also be interested in understanding more about how bit depth and sample rate interact with each other with regard to accuracy of reproduction of the musical signal, given the praise accorded by audiophiles to a "one-bit" high sample rate technology like SACD/DSD.

 

[Jud]

There is accurate information and inaccurate information. One can produce 24 bits of information using any number of means; the paper was addressing the issue of accuracy.

 

Of the eight times accuracy is addressed by the 27-page paper with regard to sample rate, in five the issue is framed in terms of bit depth. Of the other three, one mentions capacitor charging and amplifier settling. Two others simply mention accuracy without talking about an underlying reason. So I quite understandably took the paper to be saying that accuracy was an issue of bit depth.

 

Considering bit depth as something entirely separate from other considerations, which is the understanding that anyone reasonably reading the paper would take from its repeated emphasis on bit depth to the exclusion of anything else except a single passing mention of capacitor charging and amplifier settling, is what I characterized as "hogwash." I can certainly understand Dan Lavry's displeasure at reading that term from someone who has such a relatively slight understanding of the issues, as I freely admit. So now, given that you apparently agree bit depth isn't the be-all and end-all, perhaps we can characterize such a view with a less loaded term, like "incorrect," or at least "incomplete."

 

I'm still looking for a quantification of the bit depth issue at the frequencies of interest, since the only bit depths quantified in the paper were for other frequencies. How many "real" or "accurate" bits of information can one obtain at 88.2 and 96kHz sample rates, and how many at 176.4 and 192kHz?

 

[Jud]

"20 bit accuracy at 50-60Hz" is clear enough.

 

From the Lavry paper:

 

as we approach 50-60Hz we get near 24 bits

 

Not clear enough, apparently. ;-)

 

So let's accept, say, 17 or 18 at 192K. Whatever, it aint 24, like many manufacturers claim.

 

Except this assumes two things, one of which we've learned is wrong from Lavry/Brad Johnson, and the other of which is not at all certain.

 

The thing we've learned is wrong is that the bit depth numbers in the 2004 papers are not permanent truths, so if 24 bits was possible near 50 or 60Hz in 2004, perhaps it's 32 by now. I have no idea - all Brad/Lavry said is that we were incorrect in taking the remark in the paper about "permanent" to reference the particular bit depth numbers, rather than the speed/accuracy tradeoff.

 

The thing that's not at all certain (to me, anyway) is the relationship between bit depth numbers at 100mHz, 1mHz, 50 or 60Hz, 192kHz, and 96kHz. Is it linear? Logarithmic? Some other curve? The bit depth numbers given in the Lavry paper don't show a linear relationship (i.e., constant bit rate).

 

At the end of the day, it is whether the differences in bit depths among the frequencies of interest (88.2, 96, 176.4, 192) have any resulting audible significance that is the important question. No one's yet provided the bit depth figures, let alone placed them in a context that would allow us amateurs to understand whether any differences would have audible significance.

 

[Jud]

Though I can't speak for him, I think it safe to assume Keith J has also experienced a good variety of situations and has also found a clear, unambiguous pattern has emerged.

 

In fact KOJ designed a CD player for Spectral Audio very recently, and the differences between 16/44.1 and higher resolutions were a particular focus.

 

PeterSt, I (believe I) understand what you are saying about this, and am curious to listen for myself some time.

 

 

[Jud]

You keep saying Lavry says 24 bit accuracy at 192kHz is impossible. First, Lavry gave no bit depth at 192 in his 2004 paper. Second, he's already noted in this thread that whatever the limits in 2004, they've changed by now.

 

If you disagree, I'd simply ask you to provide Lavry's current bit depth number at 192. Adding 88.2 or 44.1 for comparison purposes would be nice too.

 

[Jud]

In the end, it is simply about statistics, instead of having two samples for 20 kHz tone you have more than eight.

 

This. ^^

 

I've been going on a little about bit depth because it was mentioned so much in the Lavry paper. But bit depth is only part of the story in converting analog to digital and back. I wish I knew more of the mathematics involved, but will take Miska's word in this case, since plainly material recorded at very low bit depths (SACD/DSD) can be used to produce what is accepted as audiophile-quality sound.

 

(I'd be curious, Miska, about what the calculations are that you did to determine 8x sample rates are reasonably sufficient.)

 

 

 

[Jud]

I will be receiving 200K euros for pension annually, a secretary and a car, and a place to put the secretary in.

 

Put the secretary in the car and sell his/her place. ;-)

 

[Jud]

And the cheap binoculars, advertised as 'See fifteen miles'. I can see the Andromeda Galaxy, two and a half million light years away, without any binoculars at all.

 

Reminds me very much of my friend who saw an ad for a sort of umbrella to hang above a satellite dish, advertised as being able to combat the effect of "rain fade" on signal reception, who said to me, "I wonder what they do about the other 28,000 miles?"

 

[Jud]

Only the cigar fails to be there.

 

Was it Freud or Marx (Groucho) who said "Sometimes a cigar is just a cigar"?

 

[Jud]

I agree, with one addition that seems key to me.

 

My limited understanding of what I've read here is that ~20 bits will be the practical limit due to thermal and various other forms of noise at 192kHz, *but also at the 2x rates*, so this would not be a reason to prefer lower sample rates.