[bhobba]

Is that simply because correct timing of the modes <= 22 kHz is more important than the additional content from 22 kHz -> 96 kHz that you'd get from a 192 kHz PCM?

It's I think timing. It's been known for a long time the higher the sampling rate the better it sounds and I have read you can even go into the MHz region before it stops sounding better but 768k is very very good,

But here is the rub - have a look at the MQA technical document:
http://www.aes.org/e-lib/browse.cfm?elib=17501

Notice figure 10. There is virtually nothing above the noise level beyond 48k. You can capture everything if you simply transmit at 96k, and above 24k things are 72db down which is a fairly low level so for most practical purposes red-book is fine - 96k will sound better but not by a lot - just as Rob found. This has also been my listening experience.

Just what sauce does Rob add? He has a very accurate way of taking 768k and down-sampling it to 96k, and a very accurate way of reconstructing it to 768k - much more accurate than the usual methods. Since for virtually all recordings there is nothing above the noise floor greater than 48k its an accurate reproduction of the 768k recording. MQA does something along the same lines but is deliberately not as accurate due to issues with whats called aliasing which I will not go into, because it tries tricks to exceed the 48k limit. Interestingly what they found is a very very small fraction of recordings, something like 1 in 1000, has any material above 96k and is still rather close to the noise floor so you have to question why bother capturing it? I will let others discuss the merits or otherwise of MQA but what we now know is it only affects a very small number of recordings so its like discussing angels dancing in a pinhead.

The bottom line Robs very accurate up-sampling will restore virtually all the musical information at 96k and at red-book pretty much all of it - what it doesn't reproduce is not much - audible - yes - but not greatly so.

I now have had a chance to hear my M-Scaler and TT2 for the first time and was, along with 2 other experienced Audiophiles, simply just sitting there amazed. But I will do another post about that.

Thanks
Bill

 

[bhobba]

What additional content? Rob has already stated that he cannot hear above 15KHz?

As I went into above for some reason if you increase the sampling rate to frequencies even in the Mhz region you can still hear differences, even though you can, in my case, hear nothing above 12k - if Rob can hear 15k he is doing pretty good. The conjectured reason is timing - evidently the ear can detect timing differences as small as something like 7us with 3us considered prudent.

Thanks
Bill

 

[csglinux]

It's I think timing.

I think so too, but would love to hear Rob's thoughts on exactly how. I agree with your comments on the microsecond timing nuances.

The bottom line Robs very accurate up-sampling will restore virtually all the musical information at 96k and at red-book pretty much all of it

That part I cannot agree with. That cannot be so. At least not in this universe. Here's an example to illustrate. I created a picture, which I then down-sampled as follows:

Let me know (using whatever proprietary up-sampling algorithm you like) what the original image was. I have a large cash PayPal transfer awaiting the person that gets this right

 

[514077]

What additional content? Rob has already stated that he cannot hear above 15KHz?

But it's not just those HFs we can't hear that matter. Is it not the transient timings that improve with higher sampling rates...at least in non-RW DACs?

 

[514077]

I think so too, but would love to hear Rob's thoughts on exactly how. I agree with your comments on the microsecond timing nuances.


That part I cannot agree with. That cannot be so. At least not in this universe. Here's an example to illustrate. I created a picture, which I then down-sampled as follows:



Let me know (using whatever proprietary up-sampling algorithm you like) what the original image was. I have a large cash PayPal transfer awaiting the person that gets this right

Yes, you can't restore what's lost. Ro b's DACs are the best in my opinion, but, they're not miracle workers.

 

[uzi2]

To all who quoted my post, I'm well aware that Rob detailed the marginal timing improvements with increased sampling rates in his post, but here was no mention of any additional content.

 

[csglinux]

To all who quoted my post, I'm well aware that Rob detailed the marginal timing improvements with increased sampling rates in his post, but here was no mention of any additional content.

Sorry, I don't follow. Where was there no mention of any additional content?
Where don't you think there is any additional content??

 

[bhobba]

That part I cannot agree with. That cannot be so. At least not in this universe. Here's an example to illustrate. I created a picture, which I then down-sampled as follows:

It is so because only in something like .1% (ie the one in 1000 I mentioned before) recordings has musical information above the noise floor beyond 48k. Truncating it at a sampling rate of 96k will do nothing because you just have noise - all you are getting rid of is noise. MQA uses tricks, some they will not disclose, to handle those very few recordings chopping off at 48khz makes a difference with. But MQA has problems with aliasing. They claim that at that high a frequency it's inaudible - others disagree. Rob doesn't have that issue at all - it reconstructs the original perfectly minus the noise.

I will agree however that as recording technology gets better that .1% will likely increase and we may have to go to higher transmission frequencies. But even then 96k down-sampling will still sound good.

Thanks
Bill

 

[bhobba]

But it's not just those HFs we can't hear that matter. Is it not the transient timings that improve with higher sampling rates...at least in non-RW DACs?

Yes its timing the higher sampling rates improve, but the vast number of recordings have the musical content drowned by the noise floor beyond 48khz. Sending that is of no value at all, so reconstructing a 96k samilmg rate to 768k, as far as musical content is concerned is perfect - the only difference is you have got rid of noise.

Thanks
Bill

 

[514077]

#

It is so because only in something like .1% (ie the one in 1000 I mentioned before) recordings has musical information above the noise floor beyond 46k. Truncating it at a sampling rate of 96k will do nothing because you just have noise - all you are getting rid of is noise. MQA uses tricks, some they will not disclose, to handle those very few recordings chopping off at 48khz makes a difference with. But MQA has problems with aliasing. They claim that at that high a frequency it's inaudible - others disagree. Rob doesn't have that issue at all - it reconstructs the original perfectly minus the noise.

I will agree however that as recording technology gets better that .1% will likely increase and we may have to go to higher transmission frequencies. But even then 96k down-sampling will still sound good.

Thanks
Bill

It isn't just the high frequency content that nobody supposedly does/doesn't hear. It is the increased accuracy of the timing of the transients that can improve. 44.1K has about 22ms between samples. 192kHz about 5ms or less between. Apparently, the brain is capable of noticing these differences, as it can interpret somewhere between 10 and 7ms. I've given up giving a rat's kidney about the argument over super-oral frequencies, although some feel that there may be some useful content due to hetrodyning freqs creating subharmonic
aditives to the sound of instruments. I won't go there as I was rudely shot down before. But, I've not heard anything on here to make me alter my suppositions. In general, I stick to the "higher, the better.'

 

[csglinux]

... it reconstructs the original perfectly minus the noise.

No it doesn't. If what you're saying were true, we could replace all current streaming services with a single file containing only 2 bits - and up-sampling would perfectly reconstruct Hotel California (or whatever else you were expecting) from it.

P.S. I think the confusion here is between: a) perfect and b) 16-bit or better reconstruction of the original bandwidth limited signal, i.e., up to the Nyquist limit of the original sample rate. a!=b.

 

[bhobba]

No it doesn't. If what you're saying were true, we could replace all current streaming services with a single file containing only 2 bits - and up-sampling would perfectly reconstruct Hotel California (or whatever else you were expecting) from it.

Figure 10 and some other diagrams near it in the paper I linked to before gives the noise floor of most recordings. The quantitisation noise of 24 bits it well below the thermal noise limit set by the laws of physics. Transmitting in 24 bit format is a total waste - its value is in professional work where it allows the engineer to be sloppy until the final mixdown. 20 bits is more than good enough to handle what current technology can achieve, its just a bit above the thermal noise limit. Now using a process called dithering you can make 16 bits behave as if it had more resolution ie as if it was say 20 bits. Rob has written before, while it takes some rather sophisticated dithering, his up-sampling, while transmitted at 16 bits acts like 20 so can handle anything we currently have or likely to have in the near future. It
is of no value at all transmitting noise that drowns musical content which is why transmitting at 96/16 is good enough for most recordings.

BTW DSD does 1 bit and noise shaping.

Thanks
Bill

 

[csglinux]

Figure 10 and some other diagrams near it in the paper I linked to before gives the noise floor of most recordings. The quantitisation noise of 24 bits it well below the thermal noise limit set by the laws of physics. Transmitting in 24 bit format is a total waste - its value is in professional work where it allows the engineer to be sloppy until the final mixdown. 20 bits is more than good enough to handle what current technology can achieve, its just a bit above the thermal noise limit. Now using a process called dithering you can make 16 bits behave as if it had more resolution ie as if it was say 20 bits. Rob has written before, while it takes some rather sophisticated dithering, his up-sampling, while transmitted at 16 bits acts like 20 so can handle anything we currently have or likely to have in the near future. It
is of no value at all transmitting noise that drowns musical content which is why transmitting at 96/16 is good enough for most recordings.

BTW DSD does 1 bit and noise shaping.

Thanks
Bill

Bill, you now seem to be discussing MQA, which is a totally different issue and, arguably, a bit off-topic for this thread.

I was responding to your claim that upsampling perfectly recovers the original 96/24 or 192/24 content. That isn't correct - it cannot do that. And if you're saying it's only noise beyond a certain wavenumber cut-off and so it doesn't matter, then why bother even attempting to recover it? I think it would help if Rob jumped in and explained the exact purpose of his up-sampling. I believe it is simply to reconstruct (within 16 bit resolution) the original bandwidth limited signal. I'll eat copious quantities of humble pie if I'm wrong, but I'm fairly sure Rob isn't claiming that his upsampling takes RBCD and converts it perfectly into the original hi-res file, which is what you seem to be suggesting.

 

[Rob Watts]

Lot's of posts to reply too today!

Can you expand on why it will take a long time Rob? Is it weight other projects on the go for instance?

PS: if you can’t I fully understand Rob

Primarily other projects. I think non-engineers tend to think that designing products is just turning a crank, and out pops new designs in minutes. Unfortunately, it's not so simple - each new design typically takes a few years, with 4 or 5 hardware prototypes. I am not happy with a design until it measures acceptably (and more importantly) meets SQ expectations. But Davina has been hit with hardware issues, and problems with the decimation filters - in particular the 104MHz to 768k decimation.

At the moment this year I am in Duck mode - nothing much going on on the surface but frantic paddling under the water!

Is that simply because correct timing of the modes <= 22 kHz is more important than the additional content from 22 kHz -> 96 kHz that you'd get from a 192 kHz PCM?

Absolutely. There is no real evidence that anything above 20 kHz is audible (not quite true as HF components create intermodulation through the non-linear air - but these artifacts are in the audible bandwidth and would be captured by the microphone anyway), but we have plenty of evidence that the ear/brain is sensitive to timing.

Yes, but with with higher sampling rates you can reconstruct sharper transients, which human ears have evolved to be remarkably sensitive to.

I have to disagree there; it's not about how fast the transient edges are, but whether those transients appear at the correct time. With conventional DACs the edges are constantly moving in error, backwards and forwards, depending upon the past and future music; and it's this uncertainty in transient timing that I am trying to reduce. We know that for a fact that an ideal sinc function filter will have zero uncertainty in the timing of transients, and my quest has been to make my filters as close as possible to the ideal - until such times as increasing the accuracy gives no further benefit. And I am sure that even at 1M taps on the M scaler we are not there yet.

As I went into above for some reason if you increase the sampling rate to frequencies even in the Mhz region you can still hear differences, even though you can, in my case, hear nothing above 12k - if Rob can hear 15k he is doing pretty good. The conjectured reason is timing - evidently the ear can detect timing differences as small as something like 7us with 3us considered prudent.

Thanks
Bill

Sure the evidence is that 7-4 uS is the resolution of the timing differences from between the ears; and this is based on probing cats neurons within the interaural delay, and in the case of humans asking what level of discrimination of phase shift can change the left/right placement. But I think we are much more sensitive to timing errors than this; and my evidence for this is the WTA 2 filter. This takes over from 768kHz and filters to 256 FS - so it goes from 1.3 uS resolution to 88 nS resolution. Now this replaced a perfectly good third order IIR type filter with a WTA filter, and I was surprised how easy it was to hear the difference. The change in timing POV is very small and subtle; much less than the 7-4 uS perceptual limit that the interaural delay suggests. I have also been looking at timing within noise shapers, and have heard some very small changes when timing was accurate to a few tens of nS; so my view today is that the smallest timing error is important.

I need to stress that timing errors that are audible are non-linear ones and by this I mean when the timing constantly changes with program material - and that change maybe amplitude related, or sampling and signal related. A fixed timing error is not important - so a shift of 1us at 5 kHz say is inaudible; but a 1us shift that is constantly changing is audible, as if that change is music related then the timing error confuses the brain's ability to perceive audio.

I think so too, but would love to hear Rob's thoughts on exactly how. I agree with your comments on the microsecond timing nuances.


That part I cannot agree with. That cannot be so. At least not in this universe. Here's an example to illustrate. I created a picture, which I then down-sampled as follows:



Let me know (using whatever proprietary up-sampling algorithm you like) what the original image was. I have a large cash PayPal transfer awaiting the person that gets this right

Yes agreed; I am not in the business of re-creating new information - from a mathematical sense the Nyquist Shannon sampling theory has a strictly limited information content. What the WTA filter is about is actually maintaining the information content as perfectly as possible as it goes from sampled data back to the continuous waveform. What conventional filters do is seriously degrade the original information content...

Bill, you now seem to be discussing MQA, which is a totally different issue and, arguably, a bit off-topic for this thread.

I was responding to your claim that upsampling perfectly recovers the original 96/24 or 192/24 content. That isn't correct - it cannot do that. And if you're saying it's only noise beyond a certain wavenumber cut-off and so it doesn't matter, then why bother even attempting to recover it? I think it would help if Rob jumped in and explained the exact purpose of his up-sampling. I believe it is simply to reconstruct (within 16 bit resolution) the original bandwidth limited signal. I'll eat copious quantities of humble pie if I'm wrong, but I'm fairly sure Rob isn't claiming that his upsampling takes RBCD and converts it perfectly into the original hi-res file, which is what you seem to be suggesting.

Yes it's about reconstructing the bandwidth limited signal. And I maintain that that is all one needs to do; my view is that properly bandwidth limiting to 20kHz would be perfect SQ wise.

But what do I know? We can't know anything for certain until we perform rigorous SQ tests. And that gets me nicely back to Davina, when we can actually do these tests, and will know this for certain.

 

[DjBobby]

Absolutely. There is no real evidence that anything above 20 kHz is audible (not quite true as HF components create intermodulation through the non-linear air - but these artifacts are in the audible bandwidth and would be captured by the microphone anyway), but we have plenty of evidence that the ear/brain is sensitive to timing.
I have to disagree there; it's not about how fast the transient edges are, but whether those transients appear at the correct time. With conventional DACs the edges are constantly moving in error, backwards and forwards, depending upon the past and future music; and it's this uncertainty in transient timing that I am trying to reduce. We know that for a fact that an ideal sinc function filter will have zero uncertainty in the timing of transients, and my quest has been to make my filters as close as possible to the ideal - until such times as increasing the accuracy gives no further benefit. And I am sure that even at 1M taps on the M scaler we are not there yet.

I know that you are not a fan of DSD, have read your previous posts about the DSD timing errors, but would the M-Scaler be able to "rescue" the DSD transients timing as well, or is it already lost battle in the conversion process?