Why does the soundstage sound different (often better IMHO) in high rate DSD like DSD256 Vs native Redbook to a DAC with a Chip that upsamples to ultimately do SDM conversion.

[Shadorne]

I want to discover/understand why computer based upsampling to DSD256 straight to a true 1 bit converter like in the T+A DAC 200 has better soundstage than any chip DAC based design I have ever tried. Nobody seems to know the answer yet. PC-based Redbook to high rate DSD  is so widely reported to improve soundstage that there must be a distortion mechanism in all DAC chips - is it the “equiripple” from low latency filters (cascaded 63, 31, 15 tap filters) in chip DACs that produces pre-echos in the time domain? Has anyone researched the reason for HQ player popularity? Some folks want to know - where is AES why is there no investigation by companies like Harman?

[STC]

Firstly, soundstage is the width of the two loudspeakers. Inside the width you hear object with different depth. This phantom image is created based on what reaches our ears. For the phantom images illusion to exist the speakers output the signal at different loudness and timing difference. 
 

Basically, the timing difference is within microseconds and the level difference can be from zero dB. 
 

Any improvement in soundstage by DSD should change this values. So how this values can be changed using higher sampling rate? 

[bogi]

6 hours ago, Shadorne said:

PC-based Redbook to high rate DSD  is so widely reported to improve soundstage that there must be a distortion mechanism in all DAC chips

 

I think the answer lies in less distortion and thus finer resolution of audio data which enters the D/A conversion stage when you use HQPlayer upsampling.

Just compare 2 paths, assuming possibility of direct DSD path possibility in the second one:

 

1) Audio content at 44.1k -> Player output 44.1k -> DAC chip 1st oversampling stage (filtered, for example 8x) -> DAC chip 2nd oversampling stage (unfiltered, usually ZOH, for example 16x) -> DAC chip delta sigma sigma modulator -> DAC chip D/A conversion stage -> analog signal

 

2) Audio content at 44.1k -> HQPlayer output DSD256 -> DAC chip D/A conversion stage -> analog signal

 

In 2) HQPlayer does 1 stage (for example non 2s poly-sinc filters or sinc-L) or 2 stage (-2s filters, gauss family...) filtered oversampling up to target fs, folowed by delta sigma modulation.

 

So the difference lies in what is happening before the D/A conversion stage. In the HQPlayer case it is:
- Higher precision of all calculations. HQPlayer does all in 64bit or some critical parts even in 80bit floating point precision. Higher precision allows more operations without rounding error affecting the required output resolution. DAC chips perform all calculations in limited resolution of their fixed point format.

- Higher computer power allows to run much higher number of computing operations during the limited time between 2 output samples (1/target_fs). So pure computer power allows to get higher precision result because algorithms used are not so much restricted by number and complexity of operations they can use.

- Better algorithms. Higher CPU power allows to run more advanced algorithms. Implementation quality is very important, not only computer power. So far I was experiencing only Saracon converter output giving similar quality conversion result as HQPlayer is able to provide in real time. But most probably Saracon does not provide so many filter and modulator options and AFAIK it is intended for offline conversions only. All other PCM to DSD conversion tools I tried (foobar2000, JRiver, mansr's SoX adaption, Tascam HiRes DSD editor) provide clearly lower level of output quality. Maxim Anisiutkin's foobar2000 solution provides quite nice result considering low computer resources it uses, so it is suitable as free solution for low power environments like Windows mini PCs, tablets etc.

 

All these points lead to finer resulution (lower distortion) of HQPlayer processing result in comparison with DAC chip path result. In DAC chip, the first oversampling stage runs in repeated 2x oversampling steps, where only the final 1st stage result at intermediate fs (between two stages) is filtered, but of course by simpler and lower precision algorithm. Every processing step introduces some level of signal distortion. Then the fine output resolution is further restricted by 2nd oversampling stage by 2 aspects. The first one is extremely simplified sample rate calculation (ZOH or linear interpolation) and the second one is missing filtering, resulting to audio band images present on 2nd stage output at multiples of mentioned intermediate fs. Such repeated audio band images are not present in HQPlayer DSD output. Summing up, digital signal, which enters DAC delta sigma modulator, was processed by lower quality oversampling algorithms and contains unwanted ultrasonic content, correlated with audio band, which may become source of intermodulation distortion and influence downstream equipment. That content then enters DACs delta sigma modulator. That circuit quality is of course again restricted by DAC chip hardware resources. Often it is 3rd order modulator, not in pair with HQPlayer modulators.

My understanding based on listening experience is that higher precision processing result leads to better sense of air and space, clearer instrument placement and separation, better layered soundstage (instead of flat), finer and more detailed transient presentation instead of typically hardened PCM path transient presentation, fuller and more realistic instrument timbres, better dynamics because of lower noise floor, better audible low level audibility and detail of instruments like percusion etc.

[bogi]

6 hours ago, STC said:

So how this values can be changed using higher sampling rate?

You need to assume existence of two oversampling stages in DAC chips. Oversampling as such ends at delta sigma modulator rate, which is typically about 10 MHz. So thinking like "with PCM we upsample only up to 705.6k but with PCM to DSD up to DSD256" does not consider the 2nd oversampling stage which happens in DSD chip in order to reach delta sigma modulator operating rate.

 

So the difference does not lie in higher sample rate. Both cases 1) and 2) from my previous post contain oversampling up to modulator rate, which is somewhere in MHz range. The difference then lies in digital processing quality. All what DAC chip performs before the D/A conversion stage is digital processing, so output numbers are computed from input numbers. I don't know why it is generally so hardly understandable for people that quality of digital processing can make the change. It is like when you process a picture with higher or lower quality algorithms. You get different results.

[bogi]

Correction: which happens in DSD chip -> which happens in DAC chip

[fas42]

The simple answer is that the quality of replay of digital sourced material is extremely fragile, and every tiny interference and noise factor, and slight misbehaviour in the chain, somewhere, can and mostly likely will, have an audible impact.

 

Contrary to what most believe, it has close to zero to do with how impressive various numbers of performance of bits of the digital side of things are. What really matters is how robust the chain is, to keeping under control the nasty influence of unwanted electrical cross contamination. Which of course, no-one measures ... .

[STC]

2 hours ago, bogi said:

You need to assume existence of two oversampling stages in DAC chips. Oversampling as such ends at delta sigma modulator rate, which is typically about 10 MHz. So thinking like "with PCM we upsample only up to 705.6k but with PCM to DSD up to DSD256" does not consider the 2nd oversampling stage which happens in DSD chip in order to reach delta sigma modulator operating rate.

 

So the difference does not lie in higher sample rate. Both cases 1) and 2) from my previous post contain oversampling up to modulator rate, which is somewhere in MHz range. The difference then lies in digital processing quality. All what DAC chip performs before the D/A conversion stage is digital processing, so output numbers are computed from input numbers. I don't know why it is generally so hardly understandable for people that quality of digital processing can make the change. It is like when you process a picture with higher or lower quality algorithms. You get different results.

I think I didn’t make myself clear. let me try again. 
 

1) You have a mono vocal recording. 
 

2) You convert them to stereo. 
 

3) The vocal will be in dead centre. ( In my experience, dual mono and mono converted to stereo does sound slightly different with centre position).

 

4) Now you use panning and shift the image, say 20 degrees to the left. Most panning would just involve level difference. let’s just say it is 6dB difference between the the two channels. 
 

5) The question is - How higher or lower resolution would alter that. 
 

6) It is possible the sense of depth could change due to more higher frequencies in the recording but let’s skip that for now. 
 

Thank you. 

[bogi]

5 hours ago, STC said:

I think I didn’t make myself clear. let me try again. 

 

It is like with a small object on a bitmap image. IMO it is about SNR and dynamic range what affects level of object visibility/audibility and its localization precision also in audio case.

In the case 2) with higher quality digital processing before digital signal enters the D/A conversion stage you get better SNR and higher dynamic range than in case 1). I am referring to my 1st post of this thread.

 

Then, difference in quality of digital processing between 1) and 2) results also in different level of content distortion (which is present in every digitally processed material, be it bitmap image or digital audio content).

[Shadorne]

 

12 hours ago, bogi said:

You need to assume existence of two oversampling stages in DAC chips. Oversampling as such ends at delta sigma modulator rate, which is typically about 10 MHz. So thinking like "with PCM we upsample only up to 705.6k but with PCM to DSD up to DSD256" does not consider the 2nd oversampling stage which happens in DSD chip in order to reach delta sigma modulator operating rate.

 

So the difference does not lie in higher sample rate. Both cases 1) and 2) from my previous post contain oversampling up to modulator rate, which is somewhere in MHz range. The difference then lies in digital processing quality. All what DAC chip performs before the D/A conversion stage is digital processing, so output numbers are computed from input numbers. I don't know why it is generally so hardly understandable for people that quality of digital processing can make the change. It is like when you process a picture with higher or lower quality algorithms. You get different results.

 

Good point. This difference in digital processing could be the problem. However, we are talking about devices that measure extremely well. All modern devices using processing on a chip or PC have distortion at levels that is believed to be inaudible. To affect the soundstage the distortion obviously must be audible. Since devices measure so well, it must be an extremely low level distortion in time but somehow audible - this is hard to believe but it must be true or why else would so many agree that there are audible differences between the two approaches in processing.
 

I postulate it is the pre-echo from an equiripple filter which is typically found in chip based processing. Pre-echo is an exact copy of the signal that occurs around 0.8 millisecond early but at extremely low levels - it is present whenever a filter has a constant ripple in the passband with the timing of the echo being related to the frequency of the ripple. The frequency of the ripple relates to the sample rate at which a filter is applied - so a series of filters of 63, 31 and 15 taps will create ripples at several sinusoidal frequencies corresponding to the sample rate that the filter is applied. The effect, if audible, would cause the ear brain to determine that the sound came from the direction of the speaker. The ear/brain would be confused by this pre-echo as it conflicts with timing information of the stereo image (which also has distinct time arrivals of the audio that create the illusion of position of sound within the soundstage).

 

Does anyone know if there is passband equiripple in HQplayer or Roon computer based digitally processed signal and how does this differ from typical equiripple from the simpler processing on a typical DAC chip?

 

https://en.wikipedia.org/wiki/Precedence_effect

[STC]

4 hours ago, bogi said:

visibility/audibility and its localization precision also in audio case.

 

 

It has been already well established localization is based on level difference. Higher noise cannot afford the location until the difference either increased or reduced. 
 

An example is when you hear a ticking clock in very quiet room and the same room with the air cond noise. You still will able to localize the clock because the brain decode the timing difference between the ears of the arrival of the original sound. 

[STC]

5 minutes ago, Shadorne said:

Pre-echo is an exact copy of the signal that occurs around 0.5 millisecond early but at extremely low levels whenever a filter has a constant ripple in the passband. The effect, if audible, would cause the ear brain to determine that the sound came from the direction of the speaker.

Unless this pre echo occurs only with one channel. A pre echo or post echo affects both channels simultaneous. 

[Miska]

18 minutes ago, Shadorne said:

Good point. This difference in digital processing could be the problem. However, we are talking about devices that measure extremely well. All modern devices using processing on a chip or PC have distortion at levels that is believed to be inaudible. To affect the soundstage the distortion obviously must be audible. Since devices measure so well, it must be an extremely low level distortion in time but somehow audible - this is hard to believe but it must be true or why else would so many agree that there are audible differences between the two approaches in processing.

 

Or wrong things are being measured... Or the things aren't inaudible in the end...

 

I have my own opinion on these matters...

 

18 minutes ago, Shadorne said:

I postulate it is the pre-echo from an equiripple filter which is typically found in chip based processing. Pre-echo is an exact copy of the signal that occurs around 0.5 millisecond early but at extremely low levels whenever a filter has a constant ripple in the passband.

 

Pre-echo is not related to pass-band rippleor equiripple, it is just side-effect of high order linear phase filter. Longer the filter, longer it rings. Chord being extreme example.

 

18 minutes ago, Shadorne said:

The effect, if audible, would cause the ear brain to determine that the sound came from the direction of the speaker. The ear/brain would be confused by this pre-echo as it conflicts with timing information of the stereo image (which also has distinct time arrivals of the audio that create the illusion of position of sound within the soundstage).

 

Does anyone know if there is passband equiripple in HQplayer or Roon computer based digitally processed signal and how does this differ from typical equiripple from the simpler processing on a typical DAC chip?

 

Well, this has nothing to do with pre-echo. But for example typical DAC chip filters have pass-band ripple of around 0.01 or 0.001 dB. HQPlayer filters have typically pass-band ripple below 0.000000001 dB.

 

 

But there are many factors, and only 50% of the performance comes from the digital filters. The other 50% comes from the modulator design.

 

[Shadorne]

11 minutes ago, STC said:

Unless this pre echo occurs only with one channel. A pre echo or post echo affects both channels simultaneous. 

Yes of course this pre-echo affects both channels of stereo - so it would be particularly confusing to the ear/brain if audible.

[Miska]

9 hours ago, STC said:

5) The question is - How higher or lower resolution would alter that. 

 

It is both time difference and level difference. Plus you have a very complex waveforms with lot of harmonics.

 

As the analog output signal becomes more precise, it becomes easier to distinguish the different harmonics and their timing differences.

 

[Shadorne]

6 minutes ago, Miska said:

 

Or wrong things are being measured... Or the things aren't inaudible in the end...

 

I have my own opinion on these matters...

 

 

Pre-echo is not related to pass-band rippleor equiripple, it is just side-effect of high order linear phase filter. Longer the filter, longer it rings. Chord being extreme example.

 

 

Well, this has nothing to do with pre-echo. But for example typical DAC chip filters have pass-band ripple of around 0.01 or 0.001 dB. HQPlayer filters have typically pass-band ripple below 0.000000001 dB.

 

 

But there are many factors, and only 50% of the performance comes from the digital filters. The other 50% comes from the modulator design.

 

Pre-echo or post-echo from bandpass ripple is not the same as ringing. I am not referring to ringing. Ringing is from the brick wall filter or boxcar effect or Gibbs phenomenon. I refer actually to the distinct echo from the ripples in the frequency domain of the passband.

[Miska]

8 minutes ago, Shadorne said:

Pre-echo or post-echo from bandpass ripple is not the same as ringing. I am not referring to ringing. Ringing is from the brick wall filter or boxcar effect or Gibbs phenomenon. I refer actually to the distinct echo from the ripples in the frequency domain of the passband.

 

Those are effect of Gibbs phenomenon (step- or impulse response). In HQPlayer you have number of filters as both linear- and minimum-phase variants where you can compare the effect. These are otherwise exactly the same, but minimum-phase filter has no pre-echo, just post-echo. Same pass-band ripple.

 

You are talking about pass-band ripple which is a separate parameter, just like stop-band attenuation and transition band width. Plus of course gazillion of other parameters. Pass-band ripple is just frequency response variation and doesn't cause any "echo".

 

[STC]

15 minutes ago, Miska said:

 

It is both time difference and level difference. Plus you have a very complex waveforms with lot of harmonics.

 

As the analog output signal becomes more precise, it becomes easier to distinguish the different harmonics and their timing differences.

 

For ILD and ITD, the brain expects a set of response to match what the other ear heard for localization. The experiment is easily conducted to be prove the point.  Pick or make a DSD recording with an object about 20 degrees off center. Perhaps a speech. Split the left and right channel. Down sample just one channel to 44.1 khz. Introduce a little noise that is said to be produced by the lower sampling rate. Now mix the lower sampling rate to the other untouched DSD channel. So you have one channel of 44.1 and the other a pristine DSD.  The 44.1 need to converted to DSD. Now let’s listen and see if the image changes position.

 

 

[STC]

34 minutes ago, Shadorne said:

Yes of course this pre-echo affects both channels of stereo - so it would be particularly confusing to the ear/brain if audible.

No. 5-3 is the same 6-4. The difference will always be the same. 

[Miska]

2 minutes ago, STC said:

For ILD and ITD, the brain expects a set of response to match what the other ear heard for localization. The experiment is easily conducted to be prove the point.  Pick or make a DSD recording with an object about 20 degrees off center. Perhaps a speech. Split the left and right channel. Down sample just one channel to 44.1 khz. Introduce a little noise that is said to be produced by the lower sampling rate. Now mix the lower sampling rate to the other untouched DSD channel. So you have one channel of 44.1 and the other a pristine DSD.  The 44.1 need to converted to DSD. Now let’s listen and see if the image changes position.

 

That doesn't have same effect as running DAC at lower sampling rate...

 

If you like, you can simulate the DAC errors in digital domain for one channel. But you need to do that at over 10 MHz sampling rates.

 

[Shadorne]

26 minutes ago, Miska said:

 

You are talking about pass-band ripple which is a separate parameter, just like stop-band attenuation and transition band width. Plus of course gazillion of other parameters. Pass-band ripple is just frequency response variation and doesn't cause any "echo".

 

According to Julian Dunn, equiripple in passband does cause distinct pre and post echoes. 
 

My observation of differences in soundstage (Redbook Vs pc-based DSD256 with a true 1 bit SDM DAC) has led me to conjecture that the equiripple in the passband might be the audible problem with almost all DAC chips.

Julian’s maths looks good to me - if you equate an equiripple to a sinusoid in the frequency domain then the transform is two impulse responses in the time domain. Of course the impulses are very tiny because the equiripple is so small - I would not have thought this could be audible for the post echo but it might be possible for the pre-echo. The precedence effect is not precisely understood so this is an area where measurements are not able to identify an audible effect. In any case, a tiny echo is not going to show up in any distortion measurements as the echo is a replication of any test signal exactly and precisely, at a lower level, and displaced in time.

 

https://www.nanophon.com/audio/antialia.pdf

 

[STC]

1 minute ago, Miska said:

 

That doesn't have same effect as running DAC at lower sampling rate...

 

If you like, you can simulate the DAC errors in digital domain for one channel. But you need to do that at over 10 MHz sampling rates.

 

I would appreciate that but why 10 MHz? But if you can create something like that and could hear the difference than there could be something more than ILD and ITD because I suppose you can measure exactly the ITD and ILD .

[Miska]

Just now, Shadorne said:

According to Julian Dunn, equiripple in passband does cause distinct pre and post echoes. 

 

Pre-echoes caused by linear phase filter pre-ringing is several orders of magnitude bigger.

 

Just now, Shadorne said:

My observation of differences in soundstage (Redbook Vs pc-based DSD256 with a true 1 bit SDM DAC) has led me to conjecture that the equiripple in the passband might be the audible problem with almost all DAC chips.

 

Well, HQPlayer has that covered at least, sot that it is non-issue.

 

But in my opinion, there are other bigger factors... Not limited to just digital filters.

 

[Miska]

3 minutes ago, STC said:

I would appreciate that but why 10 MHz? But if you can create something like that and could hear the difference than there could be something more than ILD and ITD because I suppose you can measure exactly the ITD and ILD .

 

Because we are discussing effects in the analog waveform reconstruction.

 

[STC]

17 minutes ago, Miska said:

 

Because we are discussing effects in the analog waveform reconstruction.

 

Okay. Anything will do as long you can demonstrate there is a shift. 

[bogi]

1 hour ago, Shadorne said:

Good point. This difference in digital processing could be the problem. However, we are talking about devices that measure extremely well. All modern devices using processing on a chip or PC have distortion at levels that is believed to be inaudible. To affect the soundstage the distortion obviously must be audible. Since devices measure so well, it must be an extremely low level distortion in time but somehow audible - this is hard to believe but it must be true or why else would so many agree that there are audible differences between the two approaches in processing.

 

There is no other difference between cases 1) and 2) than the difference in digital processing. If that difference would not exist, I would not use HQPlayer case 2). And many others with me. The case 2) is more complicated to set up by user than 1) since it requires to learn and evaluate new things and it also requires high performance computer preferably with modern nVidia card. Why would so many people use such a solution if it wouldn't have an audible effect?

I am sincerely attempting to understand things and not being a professional helps me easier to trust my ears and accept that my listening experience need not always fit to all information coming from pro world. Despite not being a pro I see that for example Amirs measurements are done too quickly and are much restricted. He is not interested to do time domain measurements except of evaluating SINAD number. He is not interested to measure above 20kHz. He ignores possible intermodulation distortion impact. He does not recognize noise impact coming from computer and power supplies to computer connected analog devices because he is not able to measure the impact which so many people are able to hear. He ignores DSD domain measurements. He ignores software upsampling topic at all. There are too many people who can distinguish 2 DACs with high SINAD numbers, impact of different power supplies, effects of USB noise treatment etc. How easy it is to simplify things so much that too simplified interpretation of restricted measurements contradicts to listening experience of so many people. From my point of view of a non pro person it creates rather comical image of some persons in a pro world. Then, of course, I evaluate which people provide information which better correlates with my hearing.

 

I only guess that many of the audible differences which we can hear but are not measured have base in time domain. I doubt it is enough well discovered what we are able to distinguish in time domain area and I also think that differences in hearing abilities between people are generally not enough considered.