As far as I know, it originally means to punish one out of ten ancient Roman soldiers, for example by beating or killing them. The DSP meaning is precisely what you wrote.
In the original valve DAC, DSD is decimated to 201.6 kHz sample rate to get it through the SRC4392 asynchronous sample rate converter. I don't know exactly how Thorp handles DSD.
I buy from Native DSD and have found their albums to be generally excellent, both in terms of the performance and the recording quality.
For what it is worth, a couple of years back I experimented with a PCM vs DSD comparison - I bought one of their samplers in both hi-res PCM and DSD formats and compaed them through HQ Player - apart from the source file and HQ Player processing everything else was identical. I felt a small preference for the native DSD playback, it seemed slighly more open and dynamic. The decoder would have been Pavel's DSC2 at that time. Obviously this is a poor comparison as it is a very small sample, entirely subjective and the differences were tiny so may be just demonstrating my psychological bias. Anyway, as a result I buy natve DSD files if they're available - yes they're huge but I take the view that disc storage is a cheap commodity.
Have you given it a try yet Acko?Got the boards already, was thinking of bypassing the second filter stage but you can also do so at your side
I will arrange to send to you asap.
Make sense here, thank you Marcel!
Yes, the analogy is pretty painful so, if there is no pcm-based SRC involved why mess around with something that already delivers quality audio in its original form? Perhaps Thorp can clarify further…
Thanks Ray, the boards look great and quality build. I am impressed by your assembly skills. I haven’t tried this as I got bogged down with other work. Just got the new ES9039Pro board assembly on my desk for verification - another DSD capable unit. so let Mark do the listening tests first and I can come round to it later.
My DAC is a DSMDAC using FPGA; the input is 96kHz/24bit PCM. I chose PCM to DSM conversion because the modulation circuit from PCM to DSM constantly evolves, and I prefer using the latest algorithms. For example, the SACD algorithm with 6th-order 1bit-64OSR significantly degrades the quality compared to the latest methods. Even with the most recent version, the conversion from PCM to DSM still needs to be improved numerically compared to DSM to PCM. The perceptual difference may vary depending on personal preference.
For example, when sampling a 12kHz sinoside at 96kHz, there are only eight independent points per cycle. The overall data would be 8x12000 (24bit) in a one-second file, but there are only eight unique points. This data can be compressed to only 8 points. However, when converting it to 1bit DSM (DSD) with 64OSR, it becomes 8x64x12000 (1bit) data, but it cannot be compressed to 8x64 points, as the first set of 8x64 points is not the same as the next set of 8x64 points.
You need to record all 8x64x12000 data to perform the correct conversion. For a two-second file, this would be 8x64x12000x2. Although occasionally, the first second's data might be repeated, it will usually be different. In the case of 1bit DSM (DSD), the smallest unit, 8x64 points, does not result in the correct conversion. You'll need a certain amount of data, say, around 8x64x1000. DSM inherently has this constraint, requiring a certain time average to achieve accurate conversion. Human ears don't respond to individual 8x64 data points but listen to their average.
This required time can be reduced by increasing the bit depth of the DSM quantizer or raising the OSR. Perceptually, 64OSR with 1bit DSM (DSD) is sufficient, but there is some degradation in the numerical conversion. That's why having the original music files in PCM format allows you to utilize conversion algorithm improvements or OSR changes.
For example, when sampling a 12kHz sinoside at 96kHz, there are only eight independent points per cycle. The overall data would be 8x12000 (24bit) in a one-second file, but there are only eight unique points. This data can be compressed to only 8 points. However, when converting it to 1bit DSM (DSD) with 64OSR, it becomes 8x64x12000 (1bit) data, but it cannot be compressed to 8x64 points, as the first set of 8x64 points is not the same as the next set of 8x64 points.
You need to record all 8x64x12000 data to perform the correct conversion. For a two-second file, this would be 8x64x12000x2. Although occasionally, the first second's data might be repeated, it will usually be different. In the case of 1bit DSM (DSD), the smallest unit, 8x64 points, does not result in the correct conversion. You'll need a certain amount of data, say, around 8x64x1000. DSM inherently has this constraint, requiring a certain time average to achieve accurate conversion. Human ears don't respond to individual 8x64 data points but listen to their average.
This required time can be reduced by increasing the bit depth of the DSM quantizer or raising the OSR. Perceptually, 64OSR with 1bit DSM (DSD) is sufficient, but there is some degradation in the numerical conversion. That's why having the original music files in PCM format allows you to utilize conversion algorithm improvements or OSR changes.
In this discussion about formats and formatting I would like to add a few comments.
As long as conversion is basically lossless, like packing native DSD into a DoP (DSD on PCM) format and consequently splitting this into left, right and clock, no loss of audio content should happen. Packing native DSD into DoP is not like oversampling and filtering.
But as 44.1/16 requires a bitstream speed of 2.116Mhz, it’s already 5.644Mhz for DSD64 and for DSD512 it’s even 45.158Mhz, more than 20 times faster then 44.1/16.
So, DSD64 requires a 176.4 PCM package, DSD128 requires a 352.8 PCM package, and DSD256 requires a 705.6 PCM package to give some idea of the huge speed differences to 44.1 PCM.
What I noticed when testing Marcel’s shiftregister Firdac in combination with the Amanero Combo384 , was a drop in S/N the higher the DSD multiplicator became.
But I also noticed this drop in S/N at increasing DSD multiples in the test’s that Archimago showed on his site for several Topping Dac’s.
I seems logical that when S/N is affected, that the whole Audio content is affected in some way.
In a lossless process where S/N goes down, the most likely reason IMO seems to be crosstalk, because of the huge transfer speeds at radio frequencies.
Could this be noticed when listening to the music ? ?
Apart from this, the S/N of a DAC isn’t the only figure of merit as long as it is good enough, a better figure doesn’t most likely improve the sound, and there are many other things that are important like a very low noise floor modulation by the signal, the absence of idle tones outside the audio band but modulating the audio signal, etc, etc. Marcel’s Solid state Dac performed exceptionally well in this respect using PWM8.
The question rises in this respect, is more better or is less more ? That’s why I performed a few tests on this audio forum to get some feeling on what seems to matters and what matters less.
Test one was using a 192/24 audio recording with Cymbals where the content went well above 20Khz, this was alternative 1.
By cleverly masking the content above 20Khz with uncorrelated noise as alternative 2 and additionally filling the space below 16bit also with uncorrelated noise, alternative 3 was created.
So just looking at the spectra, it was hard to see the original from the two alternatives.
Although auditioned by a very small group, but by people claiming to hear immediately the difference between 192/24 and 44/16, the most preferred one was alternative 3.
That confirmed my own feeling that we can’t hear anything above 20Khz or even much lower when getting older, but also content below 16bit does not seem to add anything.
So producing 44.1/16 content, but of course made with the highest possible attention, is good enough by all means, thereby implying that ca. 100dB(A) S/N is good enough and because we don’t hear anything above 20Khz, we can follow Nyquist telling that 44.1Khz must be adequate. High Res files seem to be an industries invention to ramp up sales.
The other test that I performed, this time together with Ken Newton, was to find out the importance of digital filters on the perceived sound reproduction.
For this test several 44.1/16 files where used, that where played both over 16 bit NOS Dac’s, so without any digital filtering in between.
But these files were also upsampled and filtered using software with up to several Mega point Fir filters.
Most modern Dac’s internally upsample any offered content to 192Khz, so for those Dac’s the output was always at 192/24, no matter whether 44.1/16 or 192/24 was offered.
The difference in this test was made by comparing the internal digital filter after upsampling 44.4/16 to 192/24 to the externally upsampled plus filtered 44.1/16 to 192/24.
Again performed by a small amount of contenders, it seemed that the longer the reconstruction filter, the better the sound was appreciated.
But surprisingly, the 16 bit NOS Dac’s came out just as well.
This supports the point that, under the condition that our audio equipment including speakers can handle this HF content, everything above 20Khz doesn’t seem to influence our sound perception, which is also very much the case with that DSD produces a massive amounts of HF signal, so that’s a positive thing to keep in mind.
The other thing is that digital filters seem to be very important and because of that, the consequence seems to be the lesser filters the better. That’s where DSD seems to be beneficial, at least when played without PCM conversion in between.
So, when taking all the above together:
DSD may have an edge over 44.1/16 PCM in case no additional filtering takes place.
The measured signal corruption at high DSD multiples plus the fact that we can’t hear anything above 20Khz and taking 100dB(A) S/N as adequate, there is IMO little reason to go to DSD256 or even higher.
Hans
As long as conversion is basically lossless, like packing native DSD into a DoP (DSD on PCM) format and consequently splitting this into left, right and clock, no loss of audio content should happen. Packing native DSD into DoP is not like oversampling and filtering.
But as 44.1/16 requires a bitstream speed of 2.116Mhz, it’s already 5.644Mhz for DSD64 and for DSD512 it’s even 45.158Mhz, more than 20 times faster then 44.1/16.
So, DSD64 requires a 176.4 PCM package, DSD128 requires a 352.8 PCM package, and DSD256 requires a 705.6 PCM package to give some idea of the huge speed differences to 44.1 PCM.
What I noticed when testing Marcel’s shiftregister Firdac in combination with the Amanero Combo384 , was a drop in S/N the higher the DSD multiplicator became.
But I also noticed this drop in S/N at increasing DSD multiples in the test’s that Archimago showed on his site for several Topping Dac’s.
I seems logical that when S/N is affected, that the whole Audio content is affected in some way.
In a lossless process where S/N goes down, the most likely reason IMO seems to be crosstalk, because of the huge transfer speeds at radio frequencies.
Could this be noticed when listening to the music ? ?
Apart from this, the S/N of a DAC isn’t the only figure of merit as long as it is good enough, a better figure doesn’t most likely improve the sound, and there are many other things that are important like a very low noise floor modulation by the signal, the absence of idle tones outside the audio band but modulating the audio signal, etc, etc. Marcel’s Solid state Dac performed exceptionally well in this respect using PWM8.
The question rises in this respect, is more better or is less more ? That’s why I performed a few tests on this audio forum to get some feeling on what seems to matters and what matters less.
Test one was using a 192/24 audio recording with Cymbals where the content went well above 20Khz, this was alternative 1.
By cleverly masking the content above 20Khz with uncorrelated noise as alternative 2 and additionally filling the space below 16bit also with uncorrelated noise, alternative 3 was created.
So just looking at the spectra, it was hard to see the original from the two alternatives.
Although auditioned by a very small group, but by people claiming to hear immediately the difference between 192/24 and 44/16, the most preferred one was alternative 3.
That confirmed my own feeling that we can’t hear anything above 20Khz or even much lower when getting older, but also content below 16bit does not seem to add anything.
So producing 44.1/16 content, but of course made with the highest possible attention, is good enough by all means, thereby implying that ca. 100dB(A) S/N is good enough and because we don’t hear anything above 20Khz, we can follow Nyquist telling that 44.1Khz must be adequate. High Res files seem to be an industries invention to ramp up sales.
The other test that I performed, this time together with Ken Newton, was to find out the importance of digital filters on the perceived sound reproduction.
For this test several 44.1/16 files where used, that where played both over 16 bit NOS Dac’s, so without any digital filtering in between.
But these files were also upsampled and filtered using software with up to several Mega point Fir filters.
Most modern Dac’s internally upsample any offered content to 192Khz, so for those Dac’s the output was always at 192/24, no matter whether 44.1/16 or 192/24 was offered.
The difference in this test was made by comparing the internal digital filter after upsampling 44.4/16 to 192/24 to the externally upsampled plus filtered 44.1/16 to 192/24.
Again performed by a small amount of contenders, it seemed that the longer the reconstruction filter, the better the sound was appreciated.
But surprisingly, the 16 bit NOS Dac’s came out just as well.
This supports the point that, under the condition that our audio equipment including speakers can handle this HF content, everything above 20Khz doesn’t seem to influence our sound perception, which is also very much the case with that DSD produces a massive amounts of HF signal, so that’s a positive thing to keep in mind.
The other thing is that digital filters seem to be very important and because of that, the consequence seems to be the lesser filters the better. That’s where DSD seems to be beneficial, at least when played without PCM conversion in between.
So, when taking all the above together:
DSD may have an edge over 44.1/16 PCM in case no additional filtering takes place.
The measured signal corruption at high DSD multiples plus the fact that we can’t hear anything above 20Khz and taking 100dB(A) S/N as adequate, there is IMO little reason to go to DSD256 or even higher.
Hans
Some of the reason for going to DSD256 is about stereo imaging, which we find to be audibly superior as compared to DSD64. ITD time differences between channels can correspond to much higher than 20kHz. We just don't hear such time differences as frequencies. https://en.wikipedia.org/wiki/Sound_localization
Another reason is that although we can't hear ultrasonic frequencies, those frequencies could potentially affect some parts of audio systems. If we can move the noise up farther away from the audio band then its more practical to filter with simple passive filters. Little or no need for IC op amps.
Last edited:
Hi Mark,
At all cost avoiding to get into a who’s right or wrong discussion and fully respecting your serious involvement, there are a few things to question.
You mention that WE find DSD256 superior to DSD64.
The typical Mark question would then be, who are WE and how was this test performed, true ?
The second point about ITD was discussed at length in the passed and especially Syn08 rejected this completely.
To be honest, in datastreams as fast as DSD64 and DSD256 time differences between Left and Right are way below our auditory threshold.
But let me be clear that there always may be a reason why A sounds better than B.
But finding the reason why is very hard, but if so, ITD seems very unlikely to me.
When two files DSD64 and DSD256 where generated to compare, what was their source.
There are several roads to follow, but when the DSD64 needed more filtering ……
And digital filters corrupt the sound.
Hans
At all cost avoiding to get into a who’s right or wrong discussion and fully respecting your serious involvement, there are a few things to question.
You mention that WE find DSD256 superior to DSD64.
The typical Mark question would then be, who are WE and how was this test performed, true ?
The second point about ITD was discussed at length in the passed and especially Syn08 rejected this completely.
To be honest, in datastreams as fast as DSD64 and DSD256 time differences between Left and Right are way below our auditory threshold.
But let me be clear that there always may be a reason why A sounds better than B.
But finding the reason why is very hard, but if so, ITD seems very unlikely to me.
When two files DSD64 and DSD256 where generated to compare, what was their source.
There are several roads to follow, but when the DSD64 needed more filtering ……
And digital filters corrupt the sound.
Hans
Hi Hans,
Syn08 wasn't always right.
Datastream rates are not at issue so much as phase coherence between channels.
We have a listening panel consisting of very experienced, trained listeners. NP once described one of them has still having one of the best pairs of ears in the business.
We are familiar with perceptual research methods and use our own protocol.
The reproduction system here consists of ESL speakers, dual mono block power amps, and other custom designed equipment I am not at liberty to talk about.
The DSD64 was an SACD rip. The DSD256 was made using three different modulators (for intercomparisons). Some DSD256 was made from 16/44, and some was made from 24/192.
Regarding what sounds better about A versus B, IME it is not due to one reason. There are multiple reasons and multiple differences. We aim to describe specific differences. Whether or not one version sounds better than another on average is an entirely different question. That would be a preference test, not a discrimination test.
Also as I discussed with Marcel via PM, although this is a hobby for me I am under NDA on some things.
Mark
Syn08 wasn't always right.
Datastream rates are not at issue so much as phase coherence between channels.
We have a listening panel consisting of very experienced, trained listeners. NP once described one of them has still having one of the best pairs of ears in the business.
We are familiar with perceptual research methods and use our own protocol.
The reproduction system here consists of ESL speakers, dual mono block power amps, and other custom designed equipment I am not at liberty to talk about.
The DSD64 was an SACD rip. The DSD256 was made using three different modulators (for intercomparisons). Some DSD256 was made from 16/44, and some was made from 24/192.
Regarding what sounds better about A versus B, IME it is not due to one reason. There are multiple reasons and multiple differences. We aim to describe specific differences. Whether or not one version sounds better than another on average is an entirely different question. That would be a preference test, not a discrimination test.
Also as I discussed with Marcel via PM, although this is a hobby for me I am under NDA on some things.
Mark
Last edited:
Hi Mark,
Nobody is always right, but Syn08 wasn´t exactly a dummy.
The things that you can´t disclose now, will they be released once having come to conclusions ?
The thing I agree with you is that the stereo image is a very important if not the most important reproduction aspect, but most of the time those differences in stereo images can't be measured.
I have an open mind, but in general explanations why things are sounding better or different don't hold very long until the next theory is developed.
So I prefer not to see too many explanations at all, but I'm highly interested in the outcome of serious tests such as the one you seem to be involved into.
Hans
Nobody is always right, but Syn08 wasn´t exactly a dummy.
The things that you can´t disclose now, will they be released once having come to conclusions ?
The thing I agree with you is that the stereo image is a very important if not the most important reproduction aspect, but most of the time those differences in stereo images can't be measured.
I have an open mind, but in general explanations why things are sounding better or different don't hold very long until the next theory is developed.
So I prefer not to see too many explanations at all, but I'm highly interested in the outcome of serious tests such as the one you seem to be involved into.
Hans
Eventually more details will come out. It could take some months though. However, it is possible for someone to come here for a visit and hear for themselves. Enough time and I think you will read about some of it elsewhere. Other things will remain proprietary. Not up to me.
Interesting editorial comments on “Towards an Objective Understanding of High-End Audio” in the August 2023 issue of Audio Express. A research paper on some of it at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4437822 (appears to be a free download)
As briefly mentioned in the report of Hans, he originally measured far worse DSD512 results in one channel. It turned out to be a weird kind of click noise that especially occurred at DSD512. There were strings of 16 consecutive ones coming out of the Amanero when there was a click. That is, the DAC worked fine, but there was some application software or driver or computer or Amanero issue affecting one of the two stereo channels. In the end, he just decided to measure the other channel.
When something like that happens during a comparative listening test, it could cause people to erroneously conclude that the DAC can't handle DSD512 well.
Even on the good channel, the DSD512 noise floor Hans measured with the Amanero was worse than the 27 Mbit/s (faster than DSD512) noise floor I measured without Amanero, with well-separated clock and data lines. That was still true when the exact same modulator algorithm was used for the measurements of Hans and me. I suspect data to clock crosstalk on the Amanero board, but it could also be due to some difference between the measurement set-ups.
As far as I'm concerned, the main advantages of a relatively high bit rate are that you can use quasi-multibit algorithms that can avoid idle tone issues and that you can keep the audio noise reasonably low over the feline auditory range.
Before any measurement is done, it is best to do sound checks first. I am not sure if Hans has done this already? If it doesn’t sound right then something is wrong with the setup. Ray has done listening tests with JLSounds so maybe this is a better ‘source’ than Amanero and the results were very positive.