Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[bblackwood]

Hi Bruno (and others), there is a discussion happening here in Whatever Works where many of us have heard DSD sound simply stunning vs the PCM counterpart. IME, I've had several projects now that came in for mastering that had recorded to PCM (24/96 in all cases) through very good ADC's and played back in my mastering room via very good DACs as well as recorded to the Tascam DSD recorder via it's own internal converters (played back via stock converters as well). In every case, the DSD sounded noticeably better. This is probably not a shock to you...

In the discussion linked above, one poster in particular noted that the tracks sounded closer to the original mix being recorded to DSD then decimated to PCM via software than recording directly to PCM via Lavry Gold converters. I then asked the question - do any of the modern DSD converters that output PCM do so by decimating the single bit stream in real time or are they utilizing separate audio paths for DSD and PCM? It would seem most logical to capture single bit then decimate to PCM, but I've not heard people rave about these converters being noticeably better than their PCM counterparts. Could it be that the decimation math is complicated enough that doing it 'off line' in non-real time is beneficial? Or is it simply a matter of subjective taste?

I, for one, would love to know if a product of yours (or dcs, etc) spit out PCM decimated from the DSD stream. If so, I'd love to have a unit to try out, as in my (limited) experience, recording mixes to DSD with stock converters beats recording to PCM at 2x rates via good outboard conversion.

Thoughts?

[Barry Hufker]

Brad,

You can find a great "white paper" Bruno wrote called DSD FAQ at this link.  I think it answers most if not all of your questions.  I found it fascinating anyway...

Barry


http://www.grimmaudio.com/whitepapers.htm

[bblackwood]

Thanks for the link. I think I knew most of that, though it does seem to indicate what I suspected - that the ultimate PCM recording (such as capturing the result of the analog processing chain in my mastering room) would come from a combination of the Grimm AD-1 and DD-1.

Thoughts, Bruno? Not asking you to pimp your own products, but would like your opinion on all of this...

[bruno putzeys]

The only thing I can say is that I find all of the typical sonic problems go away once you decide to do the filtering truly by the book, that is without the two main short-cuts typically taken in converters: halfband and equiripple.

If you design both the decimation and upsampling filters using windowed sinc and have the stop-band start no later than fs/2, you can make a 48kHz AD/DA chain that's as good as undetectable. Well Paul Frindle knew about that 10 years ago but sadly the simplest truths are the ones that keep getting rediscovered and forgotten.

You can improve existing record/playback chains to some extent by taking account of the 0.45fs passband and 0.55fs stopband specs typical of nearly all AD/DA chipsets. Place a sharp filter cutting off before 0.45fs before the DAC. Unfortunately, 0.45fs at 44.1kHz works out as exactly 20kHz so that's like being stuck between a rock and a hard place. Nevertheless it's an experiment worth trying: design a filter with Fpass=19500Hz, Fstop=20000, SR=90dB (less is acceptable if you're using a windowed sinc) and run any 44.1kHz audio through that. The filter is so long and the cut-off frequency so low that the ringing tails are audible as a slight glassiness, but especially on AB miked material the improvement in stereo imaging is little short of spectacular. On panpotted material the overall result may be less convincing.

Starting off with better converters is the icing on the cake of course, but I should stress that whether the converters are 1-bit  (e.g. DSD) or multibit is not fundamental to the issue. If DSD followed by off-line decimation sounds better happens when the decimator is better designed than the ones found on chip.

[Arf! Mastering]

Thanks, Bruno.  Another question raised in the Whatever Works thread is that some rather experienced mixers are preferring the DSD capture/playback from devices like the Korg MR1000 and Tascam DVRA1000 over their best 96/24 ADC-DAC chains.  What's going on there?  Are pro PCM converters costing over $3000/channel taking the filtering shortcuts referred to in your post above?

[bblackwood]

Ahh, thanks for the clarification, Bruno. It sounds as if it's a filtering issues as much as anything else...

[bruno putzeys]

Arf! Mastering wrote on Mon, 13 August 2007 14:40

Are pro PCM converters costing over $3000/channel taking the filtering shortcuts referred to in your post above?

Most are, to the best of my knowledge. There are very few converters around that use their own digital filtering.

[Arf! Mastering]

This has been quite a revelation.  Things start to make more sense now as to where theory and practice diverge.  Some converter makers seem to have been a little coy in showing their hand.  So, to reduce latency, or cost, short cuts have been taken in the filtering in "most" cases, even at the high end.  Is it correct to say that, given this fact, perceiving better performance at 192/24 than at 96/24 is defensible given that said filtering shortcuts have been taken?

[Tomas Danko]

I understand Lavry Gold uses it's very own decimation method using DSP, IIRC. Not sure if this is the case in the Blue series however.

There is indeed quite a penalty in computational power when doing 192 kHz, so it should seem common sense that developers using traditional techniques will be able to create better decimation algorithms using 96 kHz sample rate. I'm aware that the most common chips are already gathering data at a higher rate that feeds the decimation process, however.

[Arf! Mastering]

I believe Weiss also does their own DSP on board.

[bruno putzeys]

Arf! Mastering wrote on Mon, 13 August 2007 17:31

Is it correct to say that, given this fact, perceiving better performance at 192/24 than at 96/24 is defensible given that said filtering shortcuts have been taken?

It is.

Tomas Danko wrote on Mon, 13 August 2007 20:28

I understand Lavry Gold uses it's very own decimation method using DSP, IIRC.

I've not yet laid hands on one myself but the few remarks DL has dropped on his filtering indicate that he has taken pains to avoid the shortcuts. The filtering on the Gold had to be homegrown anyway since the converter itself is discrete and runs (if I recall correctly) at 2x oversampling so the analogue antialias filter produces quite a bit of phase shift that needs correcting in the digital filter.

Tomas Danko wrote on Mon, 13 August 2007 20:28

There is indeed quite a penalty in computational power when doing 192 kHz, so it should seem common sense that developers using traditional techniques will be able to create better decimation algorithms using 96 kHz sample rate.

The intelligent way is to use some of the bandwidth to relax the transition band spec. This quickly produces a net reduction in computational burden at 192kHz.

[PP]

[bruno putzeys]

"Pleasantly disturbed" it's called I believe Pigeon breeding and racing is so popular over here that it's almost impossible not to know people in this line of sports. Hopefully back to topic after this

[$a1Ty]

this may be a silly question, but you said about applying another filter, can you do that after your have recorded in the daw?

[bruno putzeys]

A lowpass filter cutting off below 0.45fs can be applied during mastering. When you're mastering for 44.1kHz that becomes a pretty long filter and its corner frequency is just audible. The ME should judge whether on balance the filter improves matters or not. If the recording consists largely of AB miked material, it's usually an improvement. If it's mostly intensity stereo (eg. panpotted) it's better to leave it alone because the 0.45fs-0.55fs aliasing problem is nearly impossible to detect in the absense of time-of-arrival cues. When you're mastering for 96kHz a slow-roll filter (fpass=20kHz, fstop=43.2kHz) is usually beneficial and otherwise innocuous.

Correcting for the inband ripple of the filters used at various places in the production chain is impractical because you'd need to keep track of all converters the signal goes through.

[Sin x/x]

Are there converters that actually fit Nyquist–Shannon sampling theorem at 44.1kHz?

[bruno putzeys]

Crystal Semiconductors used to have a few chips in their portfolio that went to zero at fs/2, but apparently they were rarely used. I presume most potential customers were put off by the fact that they started rolling off before the magical 20000.000Hz point at 44.1kHz. Speak of getting one's priorities wrong. Another explanation could be that Cirrus's own data sheets explained the difference as "non-aliasing filter" versus "filter optimized for digital audio". Clearly the sales dept had been grappling with specmanship issues.

Anyhow, one of those chips, the CS5397, found a home in Digital Audio Denmark's first converters. The company made quite a point of having this correct filtering and even coined a flashy new term to describe the issue (http://www.digitalaudio.dk/technical_papers/aid.pdf). They later changed their tune and joined the sampling rate race. The sales lit on their newer converters make no mention of using nonalias filters even though they still might be.

[$a1Ty]

so is that the main advantage of recording at sample rates above 44.1k as it moves the aliasing filter out of the audible range?

[bruno putzeys]

That's certainly the most obvious one. Somehow it seems to me like a bad solution to a problem that shouldn't have existed in the first place...

I'm saying nothing new though. The late Julian Dunn already gave a similar analysis (in 1998!) of what could make higher sampling rates sound better without having to postulate actual audibility of energy above 20kHz. http://www.nanophon.com/audio/antialia.pdf

[Tomas Danko]

$a1Ty wrote on Tue, 28 August 2007 13:57

so is that the main advantage of recording at sample rates above 44.1k as it moves the aliasing filter out of the audible range?

Doesn't upsampling and oversampling take care of this without the need for a higher sample rate?

[bruno putzeys]

If only the up/oversampling filters were designed the way they should. That was the problem.

[Tomas Danko]

Bruno Putzeys wrote on Wed, 29 August 2007 18:15

If only the up/oversampling filters were designed the way they should. That was the problem.

So once more it's down to implementation then, it seems. I still can't see the filters being a reason for higher sample rates, though. At least not in theory, that is.

[user of gear]

Hello,

   Improved latency of higher sample rates, berry berry good for overdubs and recording / programming virtual midi instrument stuff.

[Larrchild]

word

[bruno putzeys]

*scratches cranium*

[Sin x/x]

Audio Horror Stories vol 1

Bruno Putzeys wrote on Tue, 28 August 2007 04:00

Crystal Semiconductors used to have a few chips in their portfolio that went to zero at fs/2, but apparently they were rarely used. I presume most potential customers were put off by the fact that they started rolling off before the magical 20000.000Hz point at 44.1kHz. Speak of getting one's priorities wrong. Another explanation could be that Cirrus's own data sheets explained the difference as "non-aliasing filter" versus "filter optimized for digital audio". Clearly the sales dept had been grappling with specmanship issues. Anyhow, one of those chips, the CS5397, found a home in Digital Audio Denmark's first converters. The company made quite a point of having this correct filtering and even coined a flashy new term to describe the issue (http://www.digitalaudio.dk/technical_papers/aid.pdf). They later changed their tune and joined the sampling rate race. The sales lit on their newer converters make no mention of using nonalias filters even though they still might be.

Told by Bruno "Deadly Earnest" Putzeys.

[$a1Ty]

can we create a list of converters that use correct filtering

[user of gear]

Hello,

   Now it seems to me that if it is possible to oversample, then it must be possible to build a converter that simply samples at 192khz, with a properly designed filter.  If I understand this correctly, at 192khz it should be reasonably easy to start the filter high enough to be out of the way, while still allowing it to fully attentuate at the proper frequency to comply with Nyquist-Shannon's "requirement".


   I do not see why designers have chosen to go the oversampling and then decimating route, when it apparently is prone to a whole slew of other problems.  Since I am a musician and not an electrical engineer, I understand that I may not understand.


   One question I have is, honestly, how much of the "problem" is simply a result of it being very expensive to do it the right way, and how much of the "problem" is truly a result of actual technical impossibility.  


   Mr. Lavry seems to "indicate" that electronic components do not exist that can function well at 192khz, because 192khz is simply too fast for the components' settling times and so forth.  However, I do not see how that position can be reconciled with the fact that apparently all modern converters [including Mr. Lavry's] are using oversampling which is at a much higher rate than 192khz.  If the components cannot perform adequately at 192khz, they must really suck in the gigahertz range.  No?


   This brings me back to my original question.  Why not just make a PCM converter that simply samples at 192khz with the proper filters, rather than taking the polar route.  


   Everyone seems to agree that a 192khz converter with proper filter design and implementation would "satisfy the Nyquist-Shannon requirement", and would be a lossless and transparent converter.  There seems to be some difference of opinion as to whether or not the filter issues could be worked out with a 96khz sample rate.  I say it would be best to err on the side of sampling higher.  Why is everyone using oversampling instead?  Is it just because you only have to use one crystal and you can divide down to get the various pcm rates?  That does not seem like a very good reason to me.


   With regard to the previous post, I work with Pro Tools and 192khz sessions have the advantage when it comes to plug-in latencies, overdubs, and so forth.  So, even if there were no sonic advantage to 192khz over 96khz, the fact that the timing is tighter makes 192khz much more musical to work with and you end up with tracks that have a much better time feel.  Content [i.e. the feel of the music] is more important than conversion quality.


   Anyhow, I hope I am not showing my ignorance with these questions, and I hope someone can give me a straight, no-spin answer.  I would appreciate it.


   Faithful regards,


   User of gear  

[bruno putzeys]

A converter with no oversampling or upsampling would place all the burden of image rejection (=reconstruction in DAC) and antialiasing (in ADC) on analogue filters. Certainly if you want to come anywhere near the specifications that even shoddy digital filters readily attain, that's one huge analogue filter. And it's not phase linear (on the upside latency will be nearly zero).

That's when it's worth the extra trouble of making the converter run faster, use more practical analogue filters and do the really steep filtering in the digital domain.

Designing those digital filters "right" is not harder than to make compromised filters. Point is that halfband filters cut the number of calculations in half. That's why DSP designers love them. Equiripple filters further cut a significant slice off the filter's length compared to a windowed sinc filter (and reduces latency). So the economically minded designer will do that. Doing it really right, all in all, is not really much more expensive but as long as a chip manufacturer doesn't feel the incentive to incur even a little extra cost, and as long as an equipment designer doesn't feel the incentive to buy a slightly more expensive converter chip, it's not going to happen.

So, designing a good converter with all filtering in the analogue domain is much harder than doing an oversampled one, whether or not the filters are halfband equiripple or some more ideal kind.

Cutting latency in converters with minimal sonic impact can be done by trading phase linearity for latency. Cirrus Logic is currently using such filters in some of their chips. The magnitude response is more like that of a windowed filter (over most of the pass band the ripple is much smaller than specified), the pre-ringing is shorter than the post-ringing and the latency is much smaller than a linear-phase filter with the same magnitude response. So if compromises need to be made, that's one place to look. I do not buy into their argument that those filters should inherently sound better no matter what though - it's a more intelligent compromise than a shorter linear-phase filter but it's a compromise nonetheless.

[user of gear]

Hello,

   Thanks very much for the reply.  I am not quite sure I understand a couple of things, and I want to understand you properly since you are taking the time to explain this stuff.

1.)  You say, "Certainly if you want to come anywhere near the specifications that even shoddy digital filters readily attain, that's one huge analogue filter".  

   I do not want to seem ignorant but, do you mean the filter would have to actually be physically huge, or something else?

   Also, at 192khz I thought the filtering could be pretty relaxed and still work fine.  Would it still be too hard to have an analogue filter, even at 192khz?  Maybe I am talking about a converter that only does 192khz.

2.)  You say,  "That's *when* it's worth the extra trouble of making the converter run faster, use more practical analogue filters and do the really steep filtering in the digital domain".

   Did you mean to say "That's *why* it's worth...".  If not, what is your word *That* referring to?

3.)  You say, "Point is that halfband filters cut the number of calculations in half".

   Does that mean half the latency, basically?

4.)  Hypothethically, with a "money is no object" scenario, do you think it would be better to design a filter with no oversampling.

5.)  I am still unclear as to how it is possible to do the oversampling if Mr. Lavry's comments about problems with faster sampling are correct.  I mean, he is basically saying that faster sampling is less accurate, and that 96khz sampling is better than 192khz sampling for that reason.  Is oversampling different than regular sampling somehow so that it does not need to be accurate?

6.)  I do not really understand the "phase linear filter vs. non-phase linear filter" issue, so I guess I should do some research before I ask questions about all that.  Latency is definitely an issue with Pro Tools overdubbing, unless you "direct monitor", which creates other issues.  One thing I will say for the Digidesign 192 i/o is its latency is pretty low.  I have some "high end" converters here now also.  They sound lovely, but the conversion delay is longer.  I guess that is par for the course.


   Also, I see you are making DSD converters.  Are you planning on making any PCM converters?

   You mentioned Digital Audio Denmark, re: the chips with proper filtering.  FYI, I spoke to the guys at Digital Audio Denmark not too long ago.  I think they are still going with their "anti-aliasing" plan.  They are using a crystal semiconductors chip, but I think he says they are only using part of it or something like that.  They are oversampling at 5.66mhz and dividing down, as best as I understand it.  I am interested in the DXD idea, but I am trying to learn more.  

   Thanks again for your help.


   Faithful regards,


   User of gear

[bruno putzeys]

1. "Huge" as in "complicated".
Of course you could sample at 192kHz and have the filter start rolling off at 20kHz and that would make the filter practical again. But then you're already halfway towards an oversampled converter. If the information above 20kHz isn't interesting (you're filtering it away), the ADC is spewing much more data than you're really interested in.
2. When. Why. Whatever. It certainly makes a good reason to do things that way
3. Every other coefficient in a halfband is zero (except for the middle one which is exactly 1) which means you don't need to do any multiplication for those. The length of the filter is not cut in half, but the number crunching is.
4. Money no object I'd still do oversampling. The digital filter would only need to be about 30% more expensive to be perfectly done.
5. Increasing oversampling rate makes analogue filter design easier and converter design harder. So there's going to be an optimum. For Dan's converter topology (discrete ladder), the optimum works out as 2x oversampling.
6. It's a serious question. Converter manufacturers try squeezing out the last few samples worth of latency while DAWs sometimes have hundreds of samples in the FIFO. Who knows one day people will discover the joys of installing a small analogue desk in the live room for the band to make their own headphone mixes on.

In their current converters DAD use only the analogue front-end of the IC and the digital filtering is done separately. The only thing I wasn't sure of is whether they had nonaliasing or halfband. Thanks for updating me.