desirable distortion

[Fitzcaraldo215]

I am not the only one to observe that he couldn't hear the difference between direct mike feed and the mike feed D-A -> A-D with a PCM system. So i wonder if it's just stuff that those observing this can't hear or whether DSD is the inaccurate medium?

We, meaning you and I, have no idea what he did or did not hear. Personally, as I said, I have no interest in verifying it one way or the other. I merely reported what he said, which is obviously a mere anecdote, and it should be considered as such. It sounds plausible to me given my own listening experience. However, he may have his biases or imagination, just like me or anyone else, and his opinion may be valid, useless or misleading like any other anecdote.

And, some time ago, an AES paper from Germany (Detmold?) indicated scientifically from extensive DBT trials that no audible difference could reliably be heard by test subjects between DSD and PCM 176/192k (not sure which). That does not support your conclusion that DSD is a less accurate medium. But, it also is not on average, consistent with my friend's conclusion, either. However, for now, I give my friend the possible benefit of the doubt. Some people through experience learn to hear latent sonic signatures others do not. Witness our King of Audio, Amir, who is able to discern things in blind testing beyond statistical significance that others simply do not hear beyond chance. However, understood, that my friend has not demonstrated his point beyond statistical significance in DBT testing.

Personally and anecdotally from listening to many recordings, DSD is definitely a viable hirez medium. It has its proponents and detractors on various good or bad technical grounds. But, from my comments in earlier posts, I do not consider it a "must have" in playback, unlike my friend. I happily play it with excellent results using conversion to PCM and using DSP. So, I do not defend DSD, as some do, as the ultimate. It is merely a very good hirez distribution medium, but with some special requirements, limitations and difficulties in playback, ripping, etc. due to DRM and other issues.

Anyone can reach their own conclusions on this. Simply compare on any Universal player the RBCD layer of an SACD recorded and mastered in hires vs. the stereo DSD layer, carefully level matched, of course. The output levels do differ substantially, and material from analog or RBCD masters will likely reveal little discernible difference vs. hirez. And, on some players and DACS, direct DSD to analog conversion can be compared to level matched DSD-PCM-analog conversion. I have done this many times, and I personally do not find DSD to be an inaccurate medium at all. Downloads from various sites can also be used for a similar comparison.

 

[SIY]

That does not support your conclusion that DSD is a less accurate medium.

To be clear, I didn't conclude that. I raised it as one possibility.

 

[Grave]

HRA is a joke and so is DSD.

 

[Wombat]

HRA is a joke and so is DSD.

AND ........ ??????

 

[Fitzcaraldo215]

To be clear, I didn't conclude that. I raised it as one possibility.

Fair enough.

 

[Frank Dernie]

It is entirely possible I made the mistake of missing a subtle difference when I was used to an obvious difference. When I first heard the difference between the microphone feed and the A/D- D/A section of my first DAT recorder 48/16, I felt they sounded indistinguishable, but that was in comparison with reel-to-reel tape where the difference was easy to hear. Maybe there was a slight difference I didn't notice.
As far as I am concerned the digital recordings I made at 48/16 sounded exactly like the microphone feed - for better or worse
The year before last I took some of my DATs to the Scalford Hi-Fi show, they were either dummy head or crossed pair recordings with no manipulation at all. Even though few liked the music (all classical) everybody appreciated how realistic the sound was and several asked for copies.
Over 10 years ago a member of another forum whom I forget posted some samples of a recording session he attended. They recorded both simply, and with a multitude of spot mikes mixed later. For my taste the simply miked recording sample was more realistic, but has a touch of innocuous (IMO) background noise. The multi mixed version was the one released. I thought it was an excellent recording but the simple one was more pleasing to me.

 

[Jakob1863]

<snip>

And, some time ago, an AES paper from Germany (Detmold?) indicated scientifically from extensive DBT trials that no audible difference could reliably be heard by test subjects between DSD and PCM 176/192k (not sure which).

It was PCM 176,4k (1), but they found some listeners who could reliably (2) hear differences between the two formats.

They were doing 20 trial ABX tests with 110 participants doing 145 tests overall (some listeners did two or more tests) and while in the majoritiy of the tests the null-hypothesis could not be rejected, there were 4 listeners doing better with 15/17/18 and 20 correct answers. These participants used different music samples (traditional two channel stereophonic) but all were using headphones.

The majority of tests were done with multichannel setups. Unfortunately Bech/Yang let the listener choose if they want to use two channel or multichannel without making it dependent on the experience. It appeared that the majority chooses multichannel, but only 7 used a multichannel setup at home. (3)

(1) The above mentioned AES paper was just an excerpt of the whole documentation which was Bech/Yang`s diploma thesis
(2) it depends on the meaning of "reliably" as due to the time frame no replications were done back then
(3) it was a quite elaborate experiment and is a good example for the efforts needed and the role confounding factors might play, as there was a discussion about a "click" noise included

 

[andreasmaaan]

If talking 44.1 kHz rates, you normally would like to see flat response to 20 kHz. Then you need a filter starting there, that drops response to -96 dbFS or -120 dbFS by 22,050 Hz. The transition band is 20,000 to 22,050 hz as that is the range the filter works over. Transitioning from full response to no response. So any 'ringing' is confined to those frequencies. Double that to 88.2 khz and everything doubles. So 'ringing' is only happening between 40 kHz, and 44.1 kHz. Surely we can agree we aren't going to hear what is happening at more than 40 kHz.

Apologies for coming back to this late @Blumlein 88, I've just been mulling over it for a couple of days and still can't quite reconcile what you've said with my knowledge of filters (which comes only from speaker design).

I would have thought that a LPF beginning at say 20,000hz would actually create group delay throughout a large part (theoretically all) of the audible range.

Representing this statement graphically, take an arbitrary 6th order BW LPF and plot the phase response and you get:

The phase is affected well below the transition range (theoretically down to 0 if I'm not mistaken?). Is this phase response graph not just another way of representing what would manifest as "ringing" in the impulse response. Or am I missing something?

NB: I'm not trying to imply that any of this is audible in any way, just trying to further my theoretical understanding.

Cheers,
Andreas

 

[DonH56]

Pick up a text on digital filter theory... Filters can have all types of phase response and corresponding group delay functions. FIR filters do not in general correspond to classical analog filters; some IIR types do, but even there can deviate in their response.

 

[andreasmaaan]

Pick up a text on digital filter theory... Filters can have all types of phase response and corresponding group delay functions. FIR filters do not n general correspond to classical analog filters; some IIR types do, but even there can deviate in their response.

Thanks Don, that makes it clearer and I will do that reading you suggest.

In the meantime, is there a quick answer to this question: what's the lowest sample rate that allows for a FIR filter that is flat in both amplitude and phase within the audible band? And what type of FIR filter is it?

 

[SIY]

.

In the meantime, is there a quick answer to this question: what's the lowest sample rate that allows for a FIR filter that is flat in both amplitude and phase within the audible band? And what type of FIR filter is it?

If you define the top of the audible band as 20kHz, then anything above it for an antialiasing filter can be made flat in phase and amplitude.

edit: That means, for example, that a 44.1 kHz sample rate is more than adequate for this question.

 

[andreasmaaan]

Thanks @SIY. Yes, may as well define the top of the audio band as 20Khz.

So to clarify, what you're saying is that essentially in the digital world it's possible to create a brickwall low pass filter with both perfectly flat frequency response and perfectly flat phase response below the cutoff point - correct?

If that's the case, why do we even bother with high-res audio for playback?

 

[SIY]

Yes, that's right. And you can fix digitally any amplitude or phase errors caused by an analog filter.

And I don't, by and large. Recording is a different matter.

 

[andreasmaaan]

And now I think I'm beginning to see why pre-ringing comes into play with these filters. This must be a byproduct of the kind of processing that allows for flat phase response below cutoff, if I'm not mistaken?

 

[sergeauckland]

Thanks @SIY. Yes, may as well define the top of the audio band as 20Khz.

So to clarify, what you're saying is that essentially in the digital world it's possible to create a brickwall low pass filter with both perfectly flat frequency response and perfectly flat phase response below the cutoff point - correct?

If that's the case, why do we even bother with high-res audio for playback?

I don't. Way back when, I did some blind listening tests on ADC/DAC pairs, as a straight-wire bypass test, both in a digital stream (DAC-ADC) and in an analogue stream (ADC-DAC). None of us could reliably hear any difference either way at 44.1/16 bit. That convinced me that 44.1/16 was transparent, and so perfectly adequate as a distribution medium, and anything of greater resolution was unnecessary. There is a benefit in recording of using greater bit depth, where the recording will be manipulated, and for such recordings I use 32bit float, but I can't see any requirement for greater bit depths (or higher sampling rates) for distribution of finished recordings for home reproduction.

S.

 

[DonH56]

And now I think I'm beginning to see why pre-ringing comes into play with these filters. This must be a byproduct of the kind of processing that allows for flat phase response below cutoff, if I'm not mistaken?

Yes... And note that, whilst practically you cannot reach infinite suppression, you can come pretty durn close, but it takes a really large FIR (many taps, like hundreds of thousands) so latency and power become issues in the implementation.

Note digital filters can save (delay) samples so you can apply filter terms "before" and "after" the current sample (often termed pre-cursor and post-cursor samples) and thus ringing can occur before and after the current sample, thus you often see pre-ringing in the output. In the DSP world they are still causal due to the delays in the filter, but when I first saw them (longer ago than I care to admit) they threw me for a loop. Pre-shoot is easy to understand in an active analog filter design, but the whole "my filter rings before the impossible is there" took me a few minutes to understand. Looking at the block diagram of a digital filter helps, and you can think of a tapped analog delay line with variable weights to create the filters.

 

[andreasmaaan]

Yes... And note that, whilst practically you cannot reach infinite suppression, you can come pretty durn close, but it takes a really large FIR (many taps, like hundreds of thousands) so latency and power become issues in the implementation.

Note digital filters can save (delay) samples so you can apply filter terms "before" and "after" the current sample (often termed pre-cursor and post-cursor samples) and thus ringing can occur before and after the current sample, thus you often see pre-ringing in the output. In the DSP world they are still causal due to the delays in the filter, but when I first saw them (longer ago than I care to admit) they threw me for a loop. Pre-shoot is easy to understand in an active analog filter design, but the whole "my filter rings before the impossible is there" took me a few minutes to understand. Looking at the block diagram of a digital filter helps, and you can think of a tapped analog delay line with variable weights to create the filters.

Thanks again Don. I realise I'm still a long way behind here, but I'm gonna try to clarify anyway by trying to put it into language I understand.

Are you saying that by applying (many) identical taps (and "tap" I presume to mean something like "pole"?) both before and after a sample, the group delay introduced by each "before" tap is cancelled out by the group delay introduced by each "after" tap? And that in this way you end up with a filter of many many poles, but with linear phase (at the expanse of pre- and post-ringing in the transition band, which in any case is outside the audible band)?

And could you explain what you mean by "before and after the current sample"? Thx

 

[DonH56]

Each tap can have a different weight (value) so they are not in general identical. There is a relationship among taps, poles, and zeros, but it is not in general 1:1. Remember digital filters are usually clocked so there is (e.g.) a one-clock delay for each tap, leading to latency. Group delay, the (negative) derivative of phase, is determined from the filter design and implemented by the tap weights and is in general independent of the number of taps.

Look at Wikipedia, e.g. https://en.wikipedia.org/wiki/Finite_impulse_response and https://en.wikipedia.org/wiki/Infinite_impulse_response to get a better idea of the structure. Since the filter is a string of latched multiplers you get to define what is the "current" sample and then there are samples before and after the current (main cursor) sample.

This is one of the standard texts but targets senior/grad-level college students: https://www.amazon.com/Discrete-Tim...=2025&creative=165953&creativeASIN=0131988425 .

 

[Fitzcaraldo215]

I don't. Way back when, I did some blind listening tests on ADC/DAC pairs, as a straight-wire bypass test, both in a digital stream (DAC-ADC) and in an analogue stream (ADC-DAC). None of us could reliably hear any difference either way at 44.1/16 bit. That convinced me that 44.1/16 was transparent, and so perfectly adequate as a distribution medium, and anything of greater resolution was unnecessary. There is a benefit in recording of using greater bit depth, where the recording will be manipulated, and for such recordings I use 32bit float, but I can't see any requirement for greater bit depths (or higher sampling rates) for distribution of finished recordings for home reproduction.

S.

Ok, I will quibble a little, Serge. I agree that 44k/16 is a more than adequate distribution medium, although I personally would not describe it as "perfectly" adequate.

I also have trouble with "transparent". To me there are degrees of transparency. It is not a binary, true/false description. We may think something is transparent, until we hear something else we consider more transparent, in which case our standards shift. Personally, this has happened to me frequently in audio, usually with slight degrees of greater apparent transparency over the years, but not so much in recent years. The forward pace in audio has definitely slowed, I believe, except for some schiity backsliding, among other alarming mini trends.

As to your tests, I am unclear on their specifics. But, within whatever limits they may have, I accept them as totally credible, as I do you personally. Except, other very experienced recording engineers have done similar comparisons involving RBCD, hirez PCM or DSD vs. mic feeds, and they reach very different conclusions. Some stake their careers at substantial equipment expenditure upgrade levels in preferring recording and distributiion in hirez because they view it as sounding superior. Some of those guys are also very credible to me.

And, I have my own comparisons along with friends, hopefully all of us with discerning ears, of RBCD vs. various types of hirez at different sampling frequencies/formats from the same digital master. You and I do not agree that RBCD cannot be bettered, albeit slightly, but still noticeably and preferably. Note that native analog or RBCD recordings are not likely to reveal much difference for all the reasons cited in this thread about upsampling.

The 2016 Joshua Reiss meta analysis, carefully read, including looking carefully at the individual tests he summarized, can be looked at in various ways, depending on the reader's established viewpoint. To my mind, it demonstrates that some people, particularly those with prior training in what to listen for, can discern a difference with reasonable statistical significance with hirez, though not a preference which was not normally part of the testing. Others, of course, may read it as indicating such a small overall difference so as not to be worthwhile. And, some individual tests, excluding even the infamous, poorly conducted Meyer-Moran, do not show much discrimination of hirez for whatever reason.

https://secure.aes.org/forum/pubs/journal/?ID=591

Hirez is no panacea, and there is no slam dunk case to be made for it. In the slow evolution of audio, it might be a mere blip at substantial expense and inconvenience to many for not much improvement. But, others of us hear the improvement and value it.

 

[andreasmaaan]

Each tap can have a different weight (value) so they are not in general identical. There is a relationship among taps, poles, and zeros, but it is not in general 1:1. Remember digital filters are usually clocked so there is (e.g.) a one-clock delay for each tap, leading to latency. Group delay, the (negative) derivative of phase, is determined from the filter design and implemented by the tap weights and is in general independent of the number of taps.

Look at Wikipedia, e.g. https://en.wikipedia.org/wiki/Finite_impulse_response and https://en.wikipedia.org/wiki/Infinite_impulse_response to get a better idea of the structure. Since the filter is a string of latched multiplers you get to define what is the "current" sample and then there are samples before and after the current (main cursor) sample.

This is one of the standard texts but targets senior/grad-level college students: https://www.amazon.com/Discrete-Tim...=2025&creative=165953&creativeASIN=0131988425 .

Ok thanks Don, will obviously need to get a better grounding before the details make sense here...