DSD vs PCM

[Tony Lauck]

Yes, i do. this is why i added the "also" in my comment.

As i see it, if one format is superior to another then converting from the inferior format to the superior should result in the same SQ, consistently, all the time.
And it is easy to test.

If the superior format were perfect, your conclusion would be logical. However, none of the formats being discussed are perfect. In particular, DSD has noise issues while PCM has sampling rate related issues. These amount to different distortions and when converting between formats one gets the worst aspects of both formats. (This would be true even if one used the best possible conversion.)

In particular, although most people believe that 44/16 is lower resolution than 2822400/1, it is not possible to take a 44/16 file, convert it to 2822400/1 and then convert it back to 44/16 and get the same file as originally. It might be possible to get a perfect null on (some) 44/16 files when converting 44/16 to DXD and back to 44/16. It is definitely possible to convert 44/16 to 44/24 to 44/16 and get a perfect null, although this requires special settings (disable dithering) for the last conversion that would normally not be used if the 44/24 file had originally been a 24 bit file.

Digital audio, as a digital process, depends on non-linear behavior. This is what keeps 0's 0 and 1's 1. Unfortunately, non-linear behavior is very difficult to model mathematically and intuition associated with linear systems can lead to incorrect conclusions.

 

[Al M.]

If the superior format were perfect, your conclusion would be logical. However, none of the formats being discussed are perfect. In particular, DSD has noise issues while PCM has sampling rate related issues. These amount to different distortions and when converting between formats one gets the worst aspects of both formats. (This would be true even if one used the best possible conversion.)

In particular, although most people believe that 44/16 is lower resolution than 2822400/1, it is not possible to take a 44/16 file, convert it to 2822400/1 and then convert it back to 44/16 and get the same file as originally. It might be possible to get a perfect null on (some) 44/16 files when converting 44/16 to DXD and back to 44/16. It is definitely possible to convert 44/16 to 44/24 to 44/16 and get a perfect null, although this requires special settings (disable dithering) for the last conversion that would normally not be used if the 44/24 file had originally been a 24 bit file.

Digital audio, as a digital process, depends on non-linear behavior. This is what keeps 0's 0 and 1's 1. Unfortunately, non-linear behavior is very difficult to model mathematically and intuition associated with linear systems can lead to incorrect conclusions.

Thanks for that very interesting post.

 

[Lee]

If the superior format were perfect, your conclusion would be logical. However, none of the formats being discussed are perfect. In particular, DSD has noise issues while PCM has sampling rate related issues. These amount to different distortions and when converting between formats one gets the worst aspects of both formats. (This would be true even if one used the best possible conversion.)

In particular, although most people believe that 44/16 is lower resolution than 2822400/1, it is not possible to take a 44/16 file, convert it to 2822400/1 and then convert it back to 44/16 and get the same file as originally. It might be possible to get a perfect null on (some) 44/16 files when converting 44/16 to DXD and back to 44/16. It is definitely possible to convert 44/16 to 44/24 to 44/16 and get a perfect null, although this requires special settings (disable dithering) for the last conversion that would normally not be used if the 44/24 file had originally been a 24 bit file.

Digital audio, as a digital process, depends on non-linear behavior. This is what keeps 0's 0 and 1's 1. Unfortunately, non-linear behavior is very difficult to model mathematically and intuition associated with linear systems can lead to incorrect conclusions.

Tony, that is a conversion issue, not a definitive statement on the format's inherent performance. It's quite obvious that 16/44 has far less resolution than DSD.

And converting between DSD and PCM formats is not really an issue provided that the sample rate is high enough, say 352khz.

Also, I build non-linear math models for a living and it is actually quite easy to build non-linear models with machine learning tools.

 

[Sonus]

If the superior format were perfect, your conclusion would be logical. However, none of the formats being discussed are perfect. In particular, DSD has noise issues while PCM has sampling rate related issues. These amount to different distortions and when converting between formats one gets the worst aspects of both formats. (This would be true even if one used the best possible conversion.)

In particular, although most people believe that 44/16 is lower resolution than 2822400/1, it is not possible to take a 44/16 file, convert it to 2822400/1 and then convert it back to 44/16 and get the same file as originally. It might be possible to get a perfect null on (some) 44/16 files when converting 44/16 to DXD and back to 44/16. It is definitely possible to convert 44/16 to 44/24 to 44/16 and get a perfect null, although this requires special settings (disable dithering) for the last conversion that would normally not be used if the 44/24 file had originally been a 24 bit file.

Digital audio, as a digital process, depends on non-linear behavior. This is what keeps 0's 0 and 1's 1. Unfortunately, non-linear behavior is very difficult to model mathematically and intuition associated with linear systems can lead to incorrect conclusions.

If i read you correctly then if the sample rate is high enough, converting from DSD to PCM shouldn't be a problem, right?

 

[esldude]

If i read you correctly then if the sample rate is high enough, converting from DSD to PCM shouldn't be a problem, right?

I believe you will find DSD is simply different than PCM, and therefore you will have a conversion problem. At some point the residual effects may be inaudible, but as Tony has pointed out you won't be able to convert from one to the other and back recovering a bit perfect result. So in that sense one can never be identical to the other in the strictest sense.

 

[Tony Lauck]

Tony, that is a conversion issue, not a definitive statement on the format's inherent performance. It's quite obvious that 16/44 has far less resolution than DSD.

And converting between DSD and PCM formats is not really an issue provided that the sample rate is high enough, say 352khz.

Also, I build non-linear math models for a living and it is actually quite easy to build non-linear models with machine learning tools.

This is more than a conversion problem. It is a representation problem. The encoding used in PCM is very efficient for modelling certain waveforms, such as sine waves and DC levels. A simple example is where 65536 DC levels can be encoded with 16 bits. However, with DSD there are only 2 levels, and even if one looks to the 64 samples available one can only get 64 different DC levels in this time frame because of the "bit density" encoding, a.k.a. thermometer code. I can take a PCM file containing DC at one sample rate and bit depth and easily convert it to a different sample rate and bit depth and back and will get a similar DC result. This can not be done with DSD. This proves that, for at least some signals, 44/16 has higher resolution than DSD64.

Converting DSD to PCM and back to DSD has different issues. However, doing so provides a loss of sonic quality. This can be heard on some recordings that I have, where one track of an album was pure DSD and another track was made with the sequence DSD to DXD to DSD. I also happen to have the pure DSD version of the track in question. In both cases the second generation of DSD is noticeably inferior. I can't say how to allocate this to the conversion process or the use of the DXD format. It is known that multiple generations of 24 bit PCM can be done with little signal degradation, while even a few generations of DSD64 degrade the signal. This is the reason for going to DSD128 and DSD256.

 

[Lee]

Tony,

I think you are misrepresenting the beauty of DSD. With PCM you are constructing a waveform based on looking at height at very fast points in time.

With DSD, you are building the waveform by "riding" the wave so you only need to know whether to go up or down at very, very fast points in time.

Two entirely different ways of reproducing analog signals.

I agree with you that doing format conversions in general leads to a loss in sound quality but as I mentioned before, if done at 352khz or some very high sampling rate I am not sure it is audible.

Also, you can not equate conversion examples like you did to suggest that 16/44 has more resolution than DSD. No serious audio researcher or engineer believes that. It is demonstrably false.

 

[Tony Lauck]

Tony,

I think you are misrepresenting the beauty of DSD. With PCM you are constructing a waveform based on looking at height at very fast points in time.

With DSD, you are building the waveform by "riding" the wave so you only need to know whether to go up or down at very, very fast points in time.

Two entirely different ways of reproducing analog signals.

I agree with you that doing format conversions in general leads to a loss in sound quality but as I mentioned before, if done at 352khz or some very high sampling rate I am not sure it is audible.

Also, you can not equate conversion examples like you did to suggest that 16/44 has more resolution than DSD. No serious audio researcher or engineer believes that. It is demonstrably false.

I am not denigrating the beauty of DSD. I find it an excellent format for recording music, provided that it is used only one generation from an analog signal (unless higher speed DSD is used).

Conversely, every decade since the 1980's I have found 44/16 PCM an unacceptable format for recording music, and this takes into account advances in technology. There are the occasional CD recordings with very good sound, but this is rare. I have spent enough time doing conversions from analog and hi-res PCM to 44/16, using different parameters to know first hand how the different tradeoffs sound, and there are always tradeoffs. This is why I say, today, that 44/16 is an unacceptable format. Back in the 1980's and 1990's the situation was far worse, and most CD recordings were sonic garbage compared to what one could hear on LP or tape, let alone a live microphone feed.

Today, there are higher quality PCM formats that bypass the tradeoffs associated with encoding music into 44/16. I am willing to accept that 192/24 vs. DSD64 is a matter of personal taste. I do not feel this way when comparing 96/24 vs. DSD64. Indeed, those engineers making new digital recordings at 96/24 either suffer from inferior equipment or are slightly deaf, because otherwise they would be recording at 192/24 or higher resolution PCM formats (or in DSD). Just my opinion.

Finally, it is not the case that DSD and traditional PCM are fundamentally different encodings of music. They both belong to a single mathematical family of discrete representation of analog waveforms. Their difference is in sampling rate and bit depth resolution. Both involve "modulation" of the analog signal, hence the "Pulse Code Modulation" of PCM. Many PCM recordings also use some degree of noise shaping to attempt to get a lower noise floor, further emphasizing the continuity of the two formats. That is why I characterized DSD64 as 2822400/1. "Direct Stream Digital" is a trademark.

 

[Orb]

.....

Converting DSD to PCM and back to DSD has different issues. However, doing so provides a loss of sonic quality. This can be heard on some recordings that I have, where one track of an album was pure DSD and another track was made with the sequence DSD to DXD to DSD. I also happen to have the pure DSD version of the track in question. In both cases the second generation of DSD is noticeably inferior. I can't say how to allocate this to the conversion process or the use of the DXD format. It is known that multiple generations of 24 bit PCM can be done with little signal degradation, while even a few generations of DSD64 degrade the signal. This is the reason for going to DSD128 and DSD256.

Just curious was that done at 352.8kHz and 32-bit for the conversion between DSD and PCM or lower?
Seems that was the rate and bit depth Philips research lab came to the conclusion for a specification that would not detract from DSD, just going by what Bruno has mentioned in the past.
I appreciate the language used/reported on the Philips research/experience gives a little leeway in terms of possible technical semantics on whether it is transparent or actually null perfect.
Cheers
Orb

 

[microstrip]

(...) This is why I say, today, that 44/16 is an unacceptable format. (...)

Even just as a final consumer distribution format?

 

[Tony Lauck]

I was told that the tracks were remixed to adjust the level of the soloist, since it was deemed to be too low by the artists. The orchestral sound changed between the original pure DSD and the DSD-DXD-DSD, even at those points where the soloist wasn't singing. I don't believe that I was confused as to a different recording acoustic due to the slightly different mix. I believe the change was related to the lack of transparency of DSD-DXD-DSD vs. pure DSD. I don't have DSD editing capability, so I can't conduct my own controlled experiments, but what I heard was consistent with what engineers hear who produce these great sounding recordings in the first place, so they have credibility.

I was told that the DXD method of editing was used, not the DSD-wide method. I believe that converts to 352.8/32fp but I don't see much difference between this and 352.8/24, because the 32 bit IEEE floating point format has at least 25 bits of resolution and that the basic DSD64 input format has no more than 21 equivalent bits of resolution, even if one considers resolution only below 20 kHz.

 

[Tony Lauck]

Even just as a final consumer distribution format?

Yes, even as a final consumer distribution format. 44/16 PCM is not audibly transparent. It was and still is a mid-fi format in my opinion.

Anyone who converts analog signals to digital or down converts high resolution digital and doesn't hear the limitations is not a competent audio engineer, recordist or audiophile.

BTW, I hate the word "consumer" applied to music. Music is, or should be art, not something that is consumed.

 

[Orb]

Thanks Tony,
I have seen DXD described as both 24-bit (usually seen specified like this in terms of studio products and seems this is the defacto for it) and 32-bit, I appreciate though that you mention your own thoughts regarding the 32 and 24bit so not taking DXD bit depth for granted.
Just a shame none of those internal Philips research papers on these subjects are available externally, there may be something to what Bruno says (which should be pretty accurate as he worked at the same Philips research lab at that time) but tough to say what that whole context-scope is without the technical study to read.
Cheers
Orb

 

[Bruce B]

Thanks Tony,
I have seen DXD described as both 24-bit (usually seen specified like this in terms of studio products and seems this is the defacto for it) and 32-bit, I appreciate though that you mention your own thoughts regarding the 32 and 24bit so not taking DXD bit depth for granted.
Just a shame none of those internal Philips research papers on these subjects are available externally, there may be something to what Bruno says (which should be pretty accurate as he worked at the same Philips research lab at that time) but tough to say what that whole context-scope is without the technical study to read.
Cheers
Orb

The specification given for DXD is 352.8kHz, but no mention of bit-depth. It was my understanding that the bit depth had to be "at least" 24-bit, with workstations doing internal math at 32fp

Those who work with DSD everyday have been documented to say that they can hear when DSD has been edited with PCM.

 

[Orb]

The specification given for DXD is 352.8kHz, but no mention of bit-depth. It was my understanding that the bit depth had to be "at least" 24-bit, with workstations doing internal math at 32fp

Those who work with DSD everyday have been documented to say that they can hear when DSD has been edited with PCM.

Yeah I agree, the 24-bit from what I understand for DXD usually comes from the ADC, from what I have had followed anyway.
Bear in mind the Philips conclusion was to keep it as 32-bit it seems not just for the maths, but it could be debated I guess as the following is not precise wording, Bruno said in a Grimm Audio document.

When Philips engineers were building DSD editing tools the design criterion was to find a PCM format that would not detract from the sonic capabilities of DSD.
It was found that it was possible to convert a DSD signal to 352.8kHz/32 bit and back without incurring any audible quality loss, as long as good care was taken with the filtering and remodulation stages. It follows that using the same software it is also possible to convert from DXD to DSD and back without perceivable quality loss.
If a converter or DAW cannot convert from DXD to DSD and vice versa transparently, this means the processing is incorrectly implemented. It can not be used to “prove” that DXD is more transparent sounding than DSD.

Even if kept within the DAW it is still 32-bit PCM.
And yeah I can appreciate that this cannot be for dynamic range or outside the DAW as no equipment currently has a SNR beyond or close to 24-bits, and as I mentioned a bit earlier could be technical semantics and interpretation of transparent-detract/null perfect/etc.
But then it comes down to how transparent it needs to be in terms of possibly minuscule differences (well they should be very small it seems according to Philips, emphasis on should ).

Following Bruno over the years I do not think his views has changed that much on this, but I would assume as a Grimm Audio client and a well known studio this is something you could discuss with Bruno (and others at Grimm such as Eelco and Guido Tent.
Cheers
Orb

 

[Lee]

Even just as a final consumer distribution format?

Yes, of course.

 

[Tony Lauck]

Yeah I agree, the 24-bit from what I understand for DXD usually comes from the ADC, from what I have had followed anyway.
Bear in mind the Philips conclusion was to keep it as 32-bit it seems not just for the maths, but it could be debated I guess as the following is not precise wording, Bruno said in a Grimm Audio document.

Even if kept within the DAW it is still 32-bit PCM.
And yeah I can appreciate that this cannot be for dynamic range or outside the DAW as no equipment currently has a SNR beyond or close to 24-bits, and as I mentioned a bit earlier could be technical semantics and interpretation of transparent-detract/null perfect/etc.
But then it comes down to how transparent it needs to be in terms of possibly minuscule differences (well they should be very small it seems according to Philips, emphasis on should ).

Following Bruno over the years I do not think his views has changed that much on this, but I would assume as a Grimm Audio client and a well known studio this is something you could discuss with Bruno (and others at Grimm such as Eelco and Guido Tent.
Cheers
Orb

The problem with conversions isn't "PCM" vs "DSD" as all these formats are pulse code modulation. The problem is that after combining one system with low sampling rate with another system with low bit depth the result is the intersection of the two systems' capabilities and the union of the two systems' liabilities.

There are two problems. The first problem is the low sampling rate associated with DXD. Even a sampling rate of 352.8 or 384 kHz still creates filtering issues. These are magnified by the presence of high frequency DSD noise, which will alias if not filtered out, requiring a "stronger" filter when converting DSD to DXD than would be needed when converting analog to DSD. The second problem is the stacking up of DSD noise, which will exist even if there are two generations of DSD, e.g. analog - DSD - analog - DSD - analog. Someone with DSD ADCs and DACs can comment on how this sounds. I can't as I don't have the necessary equipment to conduct listening tests.

 

[tailspn]

Yeah I agree, the 24-bit from what I understand for DXD usually comes from the ADC, from what I have had followed anyway.

Hi Orb,

DXD is 24 bit 352KHz PCM. In all instances, it is derived from a Sigma-Delta Modulator operating either 1-bit, or multi-bit Pulse Density Modulation (PDM) as the front end of an A/D converter unit.

DXD is currently the best option for highest quality post processing sweetening, mixing, and track balance, of which the vast majority of deliverable media is produced. But it's not a A/D converted format per se, it's a processed/processing format. With availability of today's hardware, it's far better to record (track) in the highest available DSD bit rate (256fs with Merging's Horus/Hapi), and if required for post processing, convert off line one time to DXD, mix, sweeten, and balance.

That accomplished (edited master), the choice(s) of delivery format is market driven, with the best being DXD itself. Converting it back to DSD is just one more detail robbing function. Converting it to lower sampling rate PCM imposes the same decimation filtering (but to a lesser degree) than the decimation filtering necessary to convert the original PDM bit stream to DXD. The best fidelity is no conversions. This can/is done with either session analog mixing/balancing, or depending on the size of the musician group and venue purest microphone technique, then DSD recording and editing with no conversions.

 

[Orb]

The problem with conversions isn't "PCM" vs "DSD" as all these formats are pulse code modulation. The problem is that after combining one system with low sampling rate with another system with low bit depth the result is the intersection of the two systems' capabilities and the union of the two systems' liabilities.

There are two problems. The first problem is the low sampling rate associated with DXD. Even a sampling rate of 352.8 or 384 kHz still creates filtering issues. These are magnified by the presence of high frequency DSD noise, which will alias if not filtered out, requiring a "stronger" filter when converting DSD to DXD than would be needed when converting analog to DSD. The second problem is the stacking up of DSD noise, which will exist even if there are two generations of DSD, e.g. analog - DSD - analog - DSD - analog. Someone with DSD ADCs and DACs can comment on how this sounds. I can't as I don't have the necessary equipment to conduct listening tests.

Well technically DSD (in reality PDM spec) is closer to PWM than PCM; case in point as a quick example PDM MEMs microphones can be connected to the Blackfin DSP (and other comparable processors/SOC/etc) directly via the DSD interface.
Cheers
Prb

 

[tailspn]

Well technically DSD (in reality PDM spec) is closer to PWM than PCM; case in point as a quick example PDM MEMs microphones can be connected to the Blackfin DSP (and other comparable processors/SOC/etc) directly via the DSD interface.
Cheers
Prb

Yes, absolutely. And unlike PCM, both are valueless, and therefore need to be converted into a digital value based system/format to be process-able. That's the virtue of DXD, but at a cost of a perceivable spaciousness loss.