DSD downloads worth the trouble?

[Jud]

So let's look at our friends Lipshitz and Vanderkooy's paper, often cited in PCM/DSD comparisons.

 

It was submitted in 2003, published in 2004. The paper says PCM can in principle be perfectly dithered, which would allow PCM to behave "like an ideal analog system, having infinite resolution below the LSB [least significant bit], no distortion, and no noise modulation." A 1-bit system (which is how L and V define DSD) can't in principle be perfectly dithered or noise-shaped.

 

OK, everyone who believes we can have music via PCM (converted from the original sigma-delta-modulated [i.e., DSD-like] signal that came from the ADC) that has infinite resolution, no distortion, and no noise, please raise your hand.

 

Right, OK. So we've got knowledgeable folks in these forums who think PCM is great, and other knowledgeable folks in these forums who think DSD is great. I've heard terrific music from both, so I'll continue to buy and listen to both.

 

There then, I guess that's settled.

[tailspn]

OK, everyone who believes we can have music via PCM (converted from the original sigma-delta-modulated [i.e., DSD-like] signal that came from the ADC) that has infinite resolution, no distortion, and no noise, please raise your hand.

 

You focused on the problem Jud. ALL A/D converters in use today have Delta-Sigma Modulators used as the conversion stage, either multi-bit, or rarer, 1-bit PDM.

 

PCM, in and of itself is perfectly capable of expressing the sampled values within the limits of its sampling rate, and word bit depth. Its flaw is the redundant information carried from one sample time to the adjacent sample time. While this makes it a poor candidate from a transmission and storage aspect, its biggest problem is the inability of processors to process it at the sample rate speeds necessary to match the Delta-Sigma Modulator A/D converter front ends bit rates. Therefore, the multi-bit, or 1-bit PDM must be decimate filtered down to the PCM sampling rate from whatever the PDM bit rate, with some degree of sound quality degradation.

 

The acoustic music recording industry's problem is one of adequate PDM production tools, and its market size to make it worth the development investment. All the tools are currently PCM based, and all the originating formats are PDM. The answer so far is to increase the PCM sample rate to lower the fold-down ratio of PDM bit rates to PCM sample rates. That recently seems to be going in reverse with the wider availability of 128fs and 256fs DSD recording, while DXD (352.8KHz PCM) has not been increased in sample rate speed (yet).

 

None of this means squat in the pop music studio biz, for there's no low level spatial and ambiance cues, or instrument detail there to start with. It's primarily a factor in acoustic music recording and production. The good news is this situation is just at this slice of time. Much work is being done on several fronts to yield either all PCM recording and tools, and/or all PDM recording and tools.

 

Meanwhile there's those producers mixing and balancing in analog at the session, and just DSD editing in post. Just like they did back in the "golden" days of location recording.

[spdif-usb]

OK, everyone who believes we can have music via PCM (converted from the original sigma-delta-modulated [i.e., DSD-like] signal that came from the ADC) that has infinite resolution, no distortion, and no noise, please raise your hand.

That is not what the paper is saying. Correctly dithered PCM doesn't have infinite resolution. Instead, it has infinite time resolution. These are two entirely different things. Furthermore, the type of distortion that is fully eliminated in PCM by using the correct dither, and that cannot be fully eliminated in DSD, is called quantization error (i.e., the error that occurs when converting an analog, or continuous signal to a discrete one, or one that is either 1-bit or multi-bit).

 

For further clarification on this, I suggest that you read this paper, particularly, the following sections:

Digital Audio Gateways (p. 2 - 3)

Linearity (p. 3 - 4)

Precision and Dynamic Range (p. 4 - 5)

[Jud]

Correctly dithered PCM doesn't have infinite resolution. Instead, it has infinite time resolution. These are two entirely different things.

 

I agree. However, the L and V article quote I posted says "having infinite resolution below the LSB" [least significant bit]. That's a measure of dynamic range, not time. So L and V aren't talking about time resolution in the passage I quoted from their paper.

 

The L and V paper dates from 2003/2004, so it's probably more helpful to talk about modern dither/noise shaping in the DSD/SDM realm with people like Miska or Yuri Korzunov, since they are actually working with current solutions.

[spdif-usb]

You focused on the problem Jud. ALL A/D converters in use today have Delta-Sigma Modulators used as the conversion stage, either multi-bit, or rarer, 1-bit PDM.

 

PCM, in and of itself is perfectly capable of expressing the sampled values within the limits of its sampling rate, and word bit depth. Its flaw is the redundant information carried from one sample time to the adjacent sample time. While this makes it a poor candidate from a transmission and storage aspect, its biggest problem is the inability of processors to process it at the sample rate speeds necessary to match the Delta-Sigma Modulator A/D converter front ends bit rates. Therefore, the multi-bit, or 1-bit PDM must be decimate filtered down to the PCM sampling rate from whatever the PDM bit rate, with some degree of sound quality degradation.

 

The acoustic music recording industry's problem is one of adequate PDM production tools, and its market size to make it worth the development investment. All the tools are currently PCM based, and all the originating formats are PDM. The answer so far is to increase the PCM sample rate to lower the fold-down ratio of PDM bit rates to PCM sample rates. That recently seems to be going in reverse with the wider availability of 128fs and 256fs DSD recording, while DXD (352.8KHz PCM) has not been increased in sample rate speed (yet).

 

None of this means squat in the pop music studio biz, for there's no low level spatial and ambiance cues, or instrument detail there to start with. It's primarily a factor in acoustic music recording and production. The good news is this situation is just at this slice of time. Much work is being done on several fronts to yield either all PCM recording and tools, and/or all PDM recording and tools.

 

Meanwhile there's those producers mixing and balancing in analog at the session, and just DSD editing in post. Just like they did back in the "golden" days of location recording.

The redundant information carried from one sample time to the adjacent sample time does not in any way make it a poor candidate from a transmission and storage aspect. This is because you are forgetting the simple fact DSD carries a lot of ultrasonic noise, making it (DSD) a poor candidate for data compression, and a poorer candidate than PCM from a transmission and storage aspect exactly because of that.

 

Furthermore, a Delta-Sigma Modulator is inherently noise shaped. The decimation filter that you speak of can be combined with another filter that will remove most of the noise above 20 kHz. In practice, both filters will often be merged into a single filter to achieve better efficency and optimized performance.

 

This noise filtering characteristic relies heavily on oversampling, as it (heavy oversampling) not only makes it possible for the noise to be shifted higher above 20 kHz (where the noise can be more effectively removed by a filter that has a more gradual cut-off slope and that can be made to introduce less unwanted filter artifacts into the audible band because of it having a more gradual cut-off slope). Heavy oversampling also helps to reduce the noise in the band of interest (i.e., 0 Hz to 20 kHz) because the same amount of total noise power is distributed over a wider frequency band (e.g., 0 Hz to 1.563 MHz in the case of 8 x oversampling with a sample rate of 192 kS/s).

 

Finally, it goes without saying that sample accuracy can be further improved by incorporating parrallel sampling in ADC.

[Paul R]

Tech advancements can be both a blessing and a curse. I would gladly listen to DSD files on my

Astell & Kern AK240 instead of a Redbook CD that has a Dynamic Range value of 5.

 

Yep, those are true words indeed!

 

I wish a tech advancement would come along to allow me to type correctly and at speed on an iPad. (*sigh*) Very little time to keep up with the interesting conversations here recently. Quite embarrassing to type something, quck glance at it and miss so many typos. Must be a little like audio listening, I know what I said so I see what should be there... (grin)

 

Paul

[spdif-usb]

I agree. However, the L and V article quote I posted says "having infinite resolution below the LSB" [least significant bit]. That's a measure of dynamic range, not time. So L and V aren't talking about time resolution in the passage I quoted from their paper.

 

The L and V paper dates from 2003/2004, so it's probably more helpful to talk about modern dither/noise shaping in the DSD/SDM realm with people like Miska or Yuri Korzunov, since they are actually working with current solutions.

Believe me when I say they are talking about temporal resolution, or time resolution. The amplitude resolution, or quantization step size, Δs is still limited, as it is defined as the input range, R divided by 2^N, i.e., Δs = R/2^N where N is the number of bits per each sample. If the correct dither is applied to a PCM channel, the only limitation that still remains in the PCM channel itself, i.e. barring the limitations of the electronics devices, is the (dithered) noise floor.

 

I disagree that it is probably more helpful to talk about modern dither/noise shaping in the DSD/SDM realm with people like Miska or Yuri Korzunov. The maths, as presented by Lipshitz & Vanderkooy, are absolutely straightforward and have not been contested by anyone in any rational manner, and, after all, nobody has contested the mathematical validity of Pythagoras' theorem, either. :grin:

[spdif-usb]

Yep, those are true words indeed!

 

I wish a tech advancement would come along to allow me to type correctly and at speed on an iPad. (*sigh*) Very little time to keep up with the interesting conversations here recently. Quite embarrassing to type something, quck glance at it and miss so many typos. Must be a little like audio listening, I know what I said so I see what should be there... (grin)

 

Paul

Windows Phone 8.1 comes with Cortana so you can just talk instead of type. I recently got a Nokia Lumia 625 with Windows Phone 8.1 but, unfortunately, Cortana is (not yet) available in my country so I'll just have to make do with my Logitech K800 wireless keyboard for now. lol

[Jud]

Believe me when I say they are talking about temporal resolution, or time resolution. The amplitude resolution, or quantization step size, Δs is still limited, as it is defined as the input range, R divided by 2^N, i.e., Δs = R/2^N where N is the number of bits per each sample. If the correct dither is applied to a PCM channel, the only limitation that still remains in the PCM channel itself, i.e. barring the limitations of the electronics devices, is the (dithered) noise floor.

 

 

In the entire paper you linked, L and V consistently refer to dither's role in dynamic range/noise floor and always use LSB as a measurement of dynamic range/noise floor rather than time resolution. This is also consistent with every other paper and discussion of the subject I've read. But I'm always open to learning new things, so I invite you to refer me to instances of discussions of dither and LSB where these terms are used to refer to time resolution (particularly LSB).

[Paul R]

Believe me when I say they are talking about temporal resolution, or time resolution. The amplitude resolution, or quantization step size, Δs is still limited, as it is defined as the input range, R divided by 2^N, i.e., Δs = R/2^N where N is the number of bits per each sample. If the correct dither is applied to a PCM channel, the only limitation that still remains in the PCM channel itself, i.e. barring the limitations of the electronics devices, is the (dithered) noise floor.

 

I disagree that it is probably more helpful to talk about modern dither/noise shaping in the DSD/SDM realm with people like Miska or Yuri Korzunov. The maths, as presented by Lipshitz & Vanderkooy, are absolutely straightforward and have not been contested by anyone in any rational manner, and, after all, nobody has contested the mathematical validity of Pythagoras' theorem, either. :grin:

 

I just quick read this paper. It strikes me as a bit out of date, and is taking for granted that people need or want to do digital things to the recording in an 16/44.1K Redbook environment. While I am rather sure their math is correct, I am not so sure it the results are applied correctly. Take for instance, the typical DSD recording where there is little or no digital manipulation of the recording. In cases like this, the argument in this paper appears to become moot.

 

Has anyone evaluated this paper with regard to an entire playback system? Are the "noise" and "imperfections" discussed in this paper really an issue in a modern reproduction system? Or are they simply non-issues or at least issues that are buried under other larger and more significant issues?

 

I don't know those answers.

 

As to the controversy over the LSB here - um guys? I don't understand the references to anything "under the least significant bit." In DSD for example, all you have is the presence of or absence of what is effectively the LSB during a specific time slice.

 

-Paul

[tailspn]

The redundant information carried from one sample time to the adjacent sample time does not in any way make it a poor candidate from a transmission and storage aspect. This is because you are forgetting the simple fact DSD carries a lot of ultrasonic noise, making it (DSD) a poor candidate for data compression, and a poorer candidate than PCM from a transmission and storage aspect exactly because of that.

 

Not correct. DSD (1-bit two level PDM) is not a digitally expressed sampled data, where the shaped noise is just trash data content riding along with the desired audio data. The composite IS the quantization noise modulated by the signal in the frequency domain. They're not separable any more than is the reflected PCM image every sample rate multiple is in PCM. They're both filtered in the analog final stage, which in the case of DSD, is the integrator.

 

To your point about data efficiency between DSD and PCM, the size of a 192KHz 24bit 2 channel file in practice is 34Mb per minute, where as a DSD 2.82MB for the same content is 44MB/minute. More comparably in terms of sound quality, a 352.8KHz DXD 2 channel file is 68MB/min, verses 44MB/min for the DSD. Sound wise, there's an easily perceivable difference (loss of detail) converting the DSD original to DXD. If it were possible to convert, then process 2.82MHz DSD to 2.82MHZ 24 bit PCM, the file size would be 544MB/minuet, verses 44MB/minute for the DSD origin. That's all due to the redundant data from sample to sample. From a data storage/data transmission standpoint, DSD is far more efficient than PCM

[tailspn]

Finally, it goes without saying that sample accuracy can be further improved by incorporating parrallel sampling in ADC.

 

Then where are those ADC's? Are those IC manufacturers just not aware of this?

[Jud]

Then where are those ADC's? Are those IC manufacturers just not aware of this?

 

Perhaps if it goes without saying, no one told them!

[spdif-usb]

In the entire paper you linked, L and V consistently refer to dither's role in dynamic range/noise floor and always use LSB as a measurement of dynamic range/noise floor rather than time resolution. This is also consistent with every other paper and discussion of the subject I've read. But I'm always open to learning new things, so I invite you to refer me to instances of discussions of dither and LSB where these terms are used to refer to time resolution (particularly LSB).

It is explained in the section titled Digital Audio Gateways (p. 2 - 3) of the paper by Bob Stuart of Meridian Audio. Here's the link again.

 

http://www.meridian-audio.com/w_paper/Coding2.PDF

 

Specifically,

In practical terms, the resolution is limited by our ability to resolve sounds in noise.

And,

if the signal is processed incorrectly (i.e. truncated) it is true that the time resolution is limited to the sampling period divided by the number of digital levels². However, when the correct dither is used the time resolution also becomes effectively infinite.

² e.g. in CD that is represented by the reciprocal of 44100 * 64K

[spdif-usb]

Then where are those ADC's? Are those IC manufacturers just not aware of this?

Sheesh...

 

http://pdfserv.maximintegrated.com/en/an/AN989.pdf

[Jud]

It is explained in the section titled Digital Audio Gateways (p. 2 - 3) of the paper by Bob Stuart of Meridian Audio.

 

Did you happen to notice a little 4 letter word in one of the quotes you showed from the Stuart (not the L and V) article? I've bolded it below:

 

However, when the correct dither is used the time resolution also becomes effectively infinite.

 

"Also"? In addition to what? Here's the full quote to put things in context:

 

One of the great discoveries in PCM was that, by adding a small random noise (that we call dither) the truncation effect can disappear. Even more important was the realisation that there is a right sort of random noise to add, and that when the right dither is used [27], the resolution of the digital system becomes infinite. What results from a sensible digitisation or digital operation then is not signal plus a highly-correlated truncation distortion, but the signal and a benign low level hiss. In practical terms, the resolution is limited by our ability to resolve sounds in noise. Just to reinforce this, we have no problem measuring (and hearing) signals of –110dB in a well-designed 16-

bit channel.

 

Regarding temporal accuracy, (ii), if the signal is processed incorrectly (i.e. truncated) it is true that the time resolution is limited to the sampling period divided by the number of digital levels. However, when the correct dither is used the time resolution also becomes effectively infinite.

 

So no, neither L and V nor Stuart talks about time resolution to the exclusion of dynamic resolution being made infinite through dither. When these papers talk about "bits," they're referring to dynamic range and, especially when talking in terms of LSB, noise floor. Whittaker-Shannon-Nyquist then says the time resolution can be effectively infinite.

 

Please don't waste our time again with these sorts of quote-mining tangents. ("Quote-mining" means to selectively quote in order to give the impression of support for one's contentions from a source that does not in fact support them.)

[spdif-usb]

I just quick read this paper. It strikes me as a bit out of date, and is taking for granted that people need or want to do digital things to the recording in an 16/44.1K Redbook environment. While I am rather sure their math is correct, I am not so sure it the results are applied correctly. Take for instance, the typical DSD recording where there is little or no digital manipulation of the recording. In cases like this, the argument in this paper appears to become moot.

 

Has anyone evaluated this paper with regard to an entire playback system? Are the "noise" and "imperfections" discussed in this paper really an issue in a modern reproduction system? Or are they simply non-issues or at least issues that are buried under other larger and more significant issues?

 

I don't know those answers.

 

As to the controversy over the LSB here - um guys? I don't understand the references to anything "under the least significant bit." In DSD for example, all you have is the presence of or absence of what is effectively the LSB during a specific time slice.

 

-Paul

If quantization error wasn't a real-world problem, nobody would use dither.

[Jud]

Sheesh...

 

http://pdfserv.maximintegrated.com/en/an/AN989.pdf

 

Once again citing something that doesn't support the position you're citing it for. I'm going to quote a couple of what I think are very relevant sentences from this 2001 manufacturer's "tutorial," but anyone else is welcome to examine the full text to see if I'm misrepresenting what it says.

 

For very-high-speed applications, time interleaving increases the overall sampling speed of a system by

operating two or more data converters in parallel.

 

Note that audio is considered to be a very low speed application.

 

What sorts of speeds are we talking about for the equipment discussed in this tutorial?

 

The test setup suggested in Figure 3, for instance, is based on the use of two MAX1444 evaluation boards³ from Maxim. The MAX1444 offers the lowest-speed grade (40Msps) available in Maxim's new 10-bit +3.3V single-supply high-speed data-converter family.

 

So when a 40 million sample per second rate just isn't enough (and that's the slowest board for the types of high speed applications they're designed for), why, you can interleave two or more and get 80, 120, 160 million samples per second and up.

[spdif-usb]

Not correct. DSD (1-bit two level PDM) is not a digitally expressed sampled data, where the shaped noise is just trash data content riding along with the desired audio data. The composite IS the quantization noise modulated by the signal in the frequency domain. They're not separable any more than is the reflected PCM image every sample rate multiple is in PCM. They're both filtered in the analog final stage, which in the case of DSD, is the integrator.

 

To your point about data efficiency between DSD and PCM, the size of a 192KHz 24bit 2 channel file in practice is 34Mb per minute, where as a DSD 2.82MB for the same content is 44MB/minute. More comparably in terms of sound quality, a 352.8KHz DXD 2 channel file is 68MB/min, verses 44MB/min for the DSD. Sound wise, there's an easily perceivable difference (loss of detail) converting the DSD original to DXD. If it were possible to convert, then process 2.82MHz DSD to 2.82MHZ 24 bit PCM, the file size would be 544MB/minuet, verses 44MB/minute for the DSD origin. That's all due to the redundant data from sample to sample. From a data storage/data transmission standpoint, DSD is far more efficient than PCM

I never said they were separable, and I wan't referring to DXD. Instead, I was pointing out the fact the high amounts of ultrasonic noise in DSD are an important part of what's causing DST-encoded .dsf / .dff files to be loads bigger than PCM converted to .flac files.

[tailspn]

I never said they were separable, and I wan't referring to DXD. Instead, I was pointing out the fact the high amounts of ultrasonic noise in DSD are an important part of what's causing DST-encoded .dsf / .dff files to be loads bigger than PCM converted to .flac files.

 

You seem to be conveniently disregarding sound quality in your rebuttal. Of course DSD files are larger for the same content than FLAC ("lossless" compressed PCM) files of deliverable sampling rates. Why not extend that logic to 44.1KHz 16 bit FLAC, it's smaller yet. See where I'm going with this?

 

DXD (352.8KHz/24 bit PCM) is currently the closest in sound quality to a DSD original (whether derived from a muli-bit PDM or DSD origin), and yet is 35% larger, and sounds inferior (low level spatial and instument information).

 

With all due respect spdif, PCM is the wax cylinder of audio formats. It was the obvious choice at the dawn of digitized audio, and is by far the easiest to process digitally. But its days are numbered, to be replaced by far more efficient codes. Only the high-end audio industry still uses it, long abandoned by most others to my knowledge.

 

As I said earlier, other than being archaic and inefficient, it successfully reports audio signals. It's problem is that it must be converted to from a PDM source, which is where the losses originate.

[spdif-usb]

Did you happen to notice a little 4 letter word in one of the quotes you showed from the Stuart (not the L and V) article? I've bolded it below:

 

 

 

"Also"? In addition to what? Here's the full quote to put things in context:

 

 

 

So no, neither L and V nor Stuart talks about time resolution to the exclusion of dynamic resolution being made infinite through dither. When these papers talk about "bits," they're referring to dynamic range and, especially when talking in terms of LSB, noise floor. Whittaker-Shannon-Nyquist then says the time resolution can be effectively infinite.

 

Please don't waste our time again with these sorts of quote-mining tangents. ("Quote-mining" means to selectively quote in order to give the impression of support for one's contentions from a source that does not in fact support them.)

You are just being confused about the difference between "resolution of the digital system" and resolution in practical terms. The resolution of the digital system becomes infinite, but you know as well as I that, in practice, the resolution is still going to be limited by the noise floor and human ability to hear, or "resolve", subtle details below this noise floor.

[spdif-usb]

Once again citing something that doesn't support the position you're citing it for. I'm going to quote a couple of what I think are very relevant sentences from this 2001 manufacturer's "tutorial," but anyone else is welcome to examine the full text to see if I'm misrepresenting what it says.

 

 

 

Note that audio is considered to be a very low speed application.

 

What sorts of speeds are we talking about for the equipment discussed in this tutorial?

 

 

 

So when a 40 million sample per second rate just isn't enough (and that's the slowest board for the types of high speed applications they're designed for), why, you can interleave two or more and get 80, 120, 160 million samples per second and up.

High speed ADCs can still help to improve accuracy in audio DSP applications. :grin:

[Paul R]

If quantization error wasn't a real-world problem, nobody would use dither.

 

Ummm- noise shaping is used primarily to increase the dynamic range in the audio band, which it does quite well. While overall, a single rate DSD file might have only 6+dB, the audio band might have 100+dB of dynamic range.

 

In PCM, I think noise shaping serves a similar purpose, but involves several other steps, including upsampling and then decimating the signal to achieve the best sound at the desired sample rate.

 

I am far from clear on how you are equating this to quantization error, but will have a look at it, and hopefully gain better understanding when I have some time on the weekend.

 

Yours,

-Paul

[spdif-usb]

You seem to be conveniently disregarding sound quality in your rebuttal. Of course DSD files are larger for the same content than FLAC ("lossless" compressed PCM) files of deliverable sampling rates. Why not extend that logic to 44.1KHz 16 bit FLAC, it's smaller yet. See where I'm going with this?

 

DXD (352.8KHz/24 bit PCM) is currently the closest in sound quality to a DSD original (whether derived from a muli-bit PDM or DSD origin), and yet is 35% larger, and sounds inferior (low level spatial and instument information).

 

With all due respect spdif, PCM is the wax cylinder of audio formats. It was the obvious choice at the dawn of digitized audio, and is by far the easiest to process digitally. But its days are numbered, to be replaced by far more efficient codes. Only the high-end audio industry still uses it, long abandoned by most others to my knowledge.

 

As I said earlier, other than being archaic and inefficient, it successfully reports audio signals. It's problem is that it must be converted to from a PDM source, which is where the losses originate.

Actually no, I am not ignoring sound quality. The quantization error is not present in correctly dithered PCM. It is present in DSD. It is a type of distortion that is digital in nature. I don't see where you are going with this. What is so hard to understand about the fact the huge amount of ultrasonic noise in DSD is actually dead weight that just causes the DST lossless compression algorithm to produce bigger files instead of helps to improve the sound? You seem to be fixated on the fact the decimation that occurs in oversampling PCM is losing sound quality. I already explained that oversampling noise shaped PCM significantly helps to reduce noise at the ADC level. Do you honestly believe the quantization error of DSD does not lose sound quality? What about the artifacts caused by the filter that is required after the DSD DAC to remove all that ultrasonic noise in order not to burn the tweeter up in a thin cloud of smoke? The list goes on. If PCM has been abandoned by most others to "your knowledge" then, to my own knowledge, frankly, your knowledge must be pretty limited...

[Jud]

You are just being confused....

 

...to my own knowledge, frankly, your knowledge must be pretty limited...

 

I was going to add a comment, but I think these statements stand on their own.