World's first NOS 24/384 filterless DAC

[Telstar]

From off last Friday the 24/384 is a fact. So, all I wanted works now.

How did it sound?   

But it was an april's fool! wasn't it? 

[PeterSt]

Haha Telstar, sure not !

Gerard, yesterday I finished what I named "Octo ArcPrediction" (8 x) and immediately I noticed the separation and profoundness of everything calling itself a cymbal. Things were creapy already and this keeps on continuing. So, this 8x step seems to bring something special - something unusual. I don't have words for it yet and it needs a get used to.

One little though ... Today I created DoubleOcto Arc Prediction.



Yeah, why ?

[msjjr]

I would also like to reserve a place in the line for a DAC.

Thanks.

[easternlethal]

For those of us who are not engineers (but only informed enthusiasts), would someone care to explain in relatively simple language why filterless DACs are better? I have been spending a few days trawling through the web and have come across some articles (written by Dan Lavry, Kusoniki etc.) but I can't understand them for the life of me. 

I also have a Weiss Dac 2 which is 24/192 but not filterless, which my wife thinks would be hard to beat, but I am keen on understanding the NOS1 better...

Thanks for the help.

[PeterSt]

Ok ... It will be obvious that I try to answer this, which does not mean others aren't allowed to go against it, or doubt my thinkings. Allright ?

In simple wording, I hope :

First of all this is about Non Oversampling (NOS). This should be read as : The topology (design) of the DAC chips used is so, that it doesn't *require* oversampling in order to operate. Thus, the 1 bit chips (with sigma-delta design) *do* require oversampling, and this is always to be read as "heavy" oversampling (like 256 times etc.). This kind of oversampling by itself requires filtering or otherwise you will be left with noise only. Now :

A Non-Oversampling DAC does not require this filtering, and this is because it doesn't make use of the oversampling technique. It just doesn't need it. This can only count for multi bit chips, with as many bits as the output actually tells. Thus, a 24 bit NOS chip, makes use of 24 physically available bits. A sigma-delta chip makes use of 1 physical bit only, or possibly a mixture of the multi-bit principle (like 5 bits) and further based on the 1 bit principle.

So, an NOS chip does not require filtering in order to operate, hence produces sound without noise without any filter. However :

This doesn't mean NOS does not require filtering at all, which is related to the harmonic distortion coming from an actually too low sample rate. For example, a tone of 22050 Hz captured with a 44100 sample rate, will be an exact square wave. There are just not enough samples to even come close to the original sine (assumed the 22050 Hz tone is to be a sine). Next, it is clear I think that a square wave doesn't sound like a sine wave. The square carries harmonics, and what you will perceive from it are tones at lower frequencies than the original 22050Hz. In other words, the original 22050Hz sine -inaudible to most- now will be very much audible.
The way this works is not easy to explain, but what it comes down to is that not only the 22050Hz is a square when captured at 44100 Hz, but when the original 22050Hz is captured (!) it won't be captured at its exact "square" points. Thus, if coincidentally the capturing machine is a few micro seconds ahead on the time axis, nothing like a square will be captured, but merely a "mess". So, the original sine won't be a sine, it won't be a square, it will be a mess. And from this the lower frequency tones emerge.

The higher the sample rate at capturing, the better the original sine will be represented. And, at an infinit sample rate, the sine will be captured / registered as a sine again.

Since redbook (CD format) is about a registered sample rate of 44100 (times per second), it is so that everything above just under 5KHz will be "messy". Thus, looking at a scope, original sines of 5KHz and up don't show as nice sines anymore. This means distortion.

If during playback the missing samples (for a good representative of sines at the higher sample rates) are filled in, the problem has gone, as will be the distortion. However, in general there is no means known (but see later) which is able to fill in the missing samples;

Clever mathematical techniques were able to create a "filtering" means, which is able to restore the original wave form. The sine from our example. This is what -amongst others- Dan Lavry explains about. If you read that, it seems waterproof. But is it ?

The (digital, mathematical) filtering techniques as decribed by DL are based upon the "knowledge" of upcoming (or past) samples. Very simplified, if you'd capture that 22050 Hz sine many times, you will know how the waveform really should be, which is by itself comes from the frequency not being an "even" amount of the sample rate. Thus, now thinking of a frequency of 22049Hz but a sample rate of 44100, if you sample this wave many many times, in the end you will have captured every point of the wave as how it really was. Just, say, average all, and there you'll have it.
The downside of this is just that necessity of many other samples, which means your actual to be output sample is subject to others. This vaguens. And as how we speak of it, this "rings" (echos). Now, the major problem is that at measuring a single frequency this will just look (and be) how it really should. There is nothing to ring, and if it rings it rings as should. But this is no real life audio !

So ...
The only way to overcome this, is creating a "filtering" means that injects the samples right on the spot, and without the knowledge of other samples. And remember, it doesn't (didn't) exist.

Each DAC which is not filterless comes with a filter of the wrong type. This is because indeed the right type doesn't exist. Of course one with the right formula could choose to hard-code that formula in the digital filtering of the DAC, but this is a not so friendly means. I mean, the development area of a DAC (which is to be seen as hardware) is not the most flexible. And now the crux : so this is why I developed a filterless DAC (which must be NOS for it), in order to provide my own means of filtering from the playback software ...

And so it happened that I first created the (prototype of) the 24 bit NOS filterless DAC, and next created the Arc Prediction Upsampling, which just is that means of "filtering" that injects the samples at the proper places.
With a filtering DAC I'd be running behind the facts.

To complete this story : In order to get rid of the harmonic distortion because of too few samples, the DAC must be 192 (176.4) KHz at least, and it should be 24 bits because upsampling can't go without using additional bits (which is another story).
384 ( 352.8 ) KHz would be more optimal because it will shift away harmonic distortion from above 96KHz to above 176.4KHz, for those who are sure that even way in the inaudible range we still perceive things. So, indeed for over a week now I am playing with "Octo Arc Prediction".

Lastly, while the filtering hence upsampling (or heavy oversampling) usually is just part of the filtering means (read : no matter what, getting rid of the harmonic distortion because of a too low sample rate in the source material always needs upsampling), it occurred to me already long ago that Arc Prediction Upsampling not only does that (the filtering), but also brings sheer higher resolution. And, at going from 176.4 to 352.8 it even occurred to me that for this reason "things" in the hardware (DAC) couldn't cope anymore, hence the sound got worse from it. So I changed those "things", and now it is better than ever before.

Anticipating on the resolution increase not being at its end yet, I already created Double Octo Arc Prediction in XXHighEnd.
Well, you can guess what that actually leads to ...

Peter

PS: I hope I didn't bring in too much of my own "stuff", but since I'm the only one approaching everything like this, I hope it is allowed without sounding like a real commercial.

[manisandher]

Nice explanation Peter.

And so it happened that I first created the (prototype of) the 24 bit NOS filterless DAC, and next created the Arc Prediction Upsampling, which just is that means of "filtering" that injects the samples at the proper places.

I think people will be shocked when they hear the NOS1 DAC with Quad (and greater) Arc Prediction. The benefits of this means of filtering is not subtle. It's totally obvious. BUT...

... it really requires a 24bit NOS filterless DAC to appreciate it. I have an oversampling DAC in my office system and prefer not using Arc Prediction at all with it. In my main listening room, I have a 24bit ADC/DAC that I believe is non-oversampling at >176.4 KHz, and I much prefer using QAP to anything else.

But this raises an interesting issue (interesting to me, at least):

Most recordings nowadays are made with delta-sigma ADC chips. What affect, if any, will this have? I mean, is the recording 'doomed' from the outset?

Any thoughts?

Mani.

[easternlethal]

Peter - Thanks for the explanation. It is all much clearer to me now (I think). If I understand correctly, the "filtering" function now occurs at the software playback level (as opposed to the DAC level).

I have some further questions, if I may. Can you please expand on why you say:

The downside of this is just that necessity of many other samples, which means your actual to be output sample is subject to others. This vaguens. And as how we speak of it, this "rings" (echos)."

I am not sure I understand how multiple samples can mess up output.

Also if I don't hear much of an improvement between Arc Prediction / Quad mode and non Arc Prediction / non Quad mode, could it be because my DAC is filtering the output of XXhighend and making them both sound the same?

And finally I thought about Manisandher's question as well and it seems to be that the arc prediction + NOS solution is designed to replicate whatever is recorded on the CD faithfully so to the extent that the recording is not as good (whether because of delta-sigma or anything else), there will be a limitation, but an Arc Predicted / NOS DAC outputted sound should still be better than a filtered DAC output in any situation.

Admiring your quest for better and better sound.

[Telstar]

... it really requires a 24bit NOS filterless DAC to appreciate it. I have an oversampling DAC in my office system and prefer not using Arc Prediction at all with it. In my main listening room, I have a 24bit ADC/DAC that I believe is non-oversampling at >176.4 KHz, and I much prefer using QAP to anything else.

Yes, indeed I prefer straight 44.1k without any upsampling on my OS AKM DAC.

[PeterSt]

But this raises an interesting issue (interesting to me, at least):

Most recordings nowadays are made with delta-sigma ADC chips. What affect, if any, will this have? I mean, is the recording 'doomed' from the outset?

Any thoughts?

Hi there Mani,

I must honestly say that not long ago I wondered this myself but couldn't come to proper conclusions. And this still is so : I just don't know - or can't reason it out. However, maybe it starts to occur to me that it is not just record companies (like HR) that produce better recordings because of their recording capabilities, but that it is "just" the Microsonics being used. Now it becomes difficult, because the one record company would be using the same "recorder" all the time.

Funnily enough I am a step further again on sound quality, and it (as it seems to me) unveils a character of sound from ... recordings with the Microsonics. It is no coincidence that I now can recognize this (again, as it seems) because I am concentrating on HDCD recordings the past few days. Just from random recording companies. Of course this *should* unveil a character of sound, because or I'm listening to undecoded HDCD, or I'm listening to decoded HDCD (hahaha) and both are not normal redbook. Ok, turning this some other way around, at listening to random (non-HDCD) HRx recordings, they are not a bit better, they are HUGELY better. Actually there's no comparison, and I don't think this comes from special mikes or whatever.
FYI: Yes, I am using *and* HDCD decoding *and* Arc Prediction (8x).

I guess it need some digging out (by me) on how A/D works in detail, and *if* it even is able to work with multi-bit only (yes, thus far I doubt that). Also, if the Microsonics would have worked with pure multibit A/D's I really wonder which that ever would have been for 192KHz (up to 2002). Did 24/192 multibit chips ever exist ? and what would have been the famous parts out of production so they had to quit the Microsonics production ?
(if true at all, because it also can be related to MS (eh. I mean M$ here) obtaining the HDCD license just before that time).

You know, this alll *is* interesting, because when I know how things are molested during the recording stage (if at all), I may be able to restore that ...
I know, this is highly opportunistic.

Peter

[AUDIODIDAKT]

Ha Peter,

Good to see you followed up, scary huh
Was about to ask about your findings on this, but with HDCD.exe available i could only imagine it will be in the next version.
anyway, would be insanely great to find and restore studio/recording flaws, maybe impossible, who knows.
Something is going on, maybe these Pacific Microsonics are just right.

How to obtain such info about recordings ?
How many other recording (software, mastering) systems are being used elsewhere ?

after that it would be the case "How to recognize such recordings by software",


You are right about unveiling certain recording characters, because, well........ this is just the case.

maybe this is next step in the playback evolution of digital music !?
Who knows.........

[PeterSt]

Peter - Thanks for the explanation. It is all much clearer to me now (I think). If I understand correctly, the "filtering" function now occurs at the software playback level (as opposed to the DAC level).

I have some further questions, if I may. Can you please expand on why you say:

The downside of this is just that necessity of many other samples, which means your actual to be output sample is subject to others. This vaguens. And as how we speak of it, this "rings" (echos)."

I am not sure I understand how multiple samples can mess up output.

Also if I don't hear much of an improvement between Arc Prediction / Quad mode and non Arc Prediction / non Quad mode, could it be because my DAC is filtering the output of XXhighend and making them both sound the same?

And finally I thought about Manisandher's question as well and it seems to be that the arc prediction + NOS solution is designed to replicate whatever is recorded on the CD faithfully so to the extent that the recording is not as good (whether because of delta-sigma or anything else), there will be a limitation, but an Arc Predicted / NOS DAC outputted sound should still be better than a filtered DAC output in any situation.

Admiring your quest for better and better sound.

With the common filtering many other samples (following and/or preceeding) "create" the current sample to be output. So, the other samples weigh in. Something like (very simplistically) the current sample weighs in for 50%, the next for 20%, the next for 18% the next for 16%, the last for 2 % (totalling to 100%) and there is the output sample. This is far from the original sample (which isn't good either), and the main objective one (like me) could have is that none of the original sample points are in the orignal sample anymore. This, however, should be looked at as a full wave of a frequency : none of the originally captured samples comprising that wave, are at their original time-axis location anymore. This can also mean the wave has shifted in time, and creates a phase error (which happens with IIR filtering and is called "minimum phase" as in "almost nothing is good anymore, haha). So, with proper interpolation (Arc Prediction is about interpolation), all the original sample points remain, and only new ones are injected.
Hoping you are not confused even more now, this is why Arc Prediction can only do "even" (like from 44.1 to 88.2) upsampling rates because uneven (like from 44.1 to 96) requires the original sample points to be lost.

Also if I don't hear much of an improvement between Arc Prediction / Quad mode and non Arc Prediction / non Quad mode, could it be because my DAC is filtering the output of XXhighend and making them both sound the same?

This would be very true. There's a small mystery going on though : for most people with OS DACs or NOS/Filtering DACs, Arc Prediction still improves;
Although I can't reason this out for 100%, this has to be related to the smartness of the filtering algorithms, which can be highly based upon "is there anything to correct, and if not, never mind". At least for Arc Prediction it works like this, and any file already subject to Arc Prediction (might it exist) won't change a thing to the wave form anymore. Thus, IMO it depends on the filter used, and I don't know them ...

And finally I thought about Manisandher's question as well and it seems to be that the arc prediction + NOS solution is designed to replicate whatever is recorded on the CD faithfully [...]

Mwah, a little different : The point is, what is recorded on the CD is wrong always, but it is the trick to bring that back to what was recorded (the nice sine etc.).
On a side note though (and this might be what you mean), if I record the output of NOS (with or without Arc Prediction, but filterless) and digitize that again (capture by means of an A/D converter), I can very well compare the CD data with the output analogue data coming from that. There's still a huge error, but it is comparable. If I do the same with an OS/Filtering DAC, no head nor tail can be made from it. It has become COMPLETELY different data. It is even a miracle that we perceive music from it. And do note please that I created the comparising software for it (actually in XXHighEnd but disabled -> the Analysis tab), and what I do in there has not been done before (and is thus unknown to "the world"). If you'd look at the graphs coming from this, well, then you know at last and for sure that not only a zillion things are wrong with audio, but that as many things can still be improved just because of that ! Isn't that a nice future ? (in case I'd be running empty of work, haha).
In the end, although nobody using good equipment with a nice OS DAC will believe it, this is exactly what you hear/perceive from OS : actually nothing much. It may be nice sound, it may be not fatigueing, the background may be nicely black, but what are the instruments you are hearing ? It just doesn't work. And if you'd see those graphs, well, you just can see why. There is no comparison with reality, that reality being the CD data, though (and I can't emphasize this enough) also wrong. For fun, do this :

Draw one cycle of a sine on a paper. The line starts at 0V, the first peak is at 2V, next it crosses 0V and dips into -2V and it ends at 0V again. So, this is one cycle of a frequency. Now assume a samplerate that catches this sine with 4 samples, and divide those samples nicely on the time line (horizontally). Make a sample where the sine starts (0V) make one at the first peak (2V), make one where it crosses 0, make one at -2V, and lastly make one where it ends, which is the same as the start of the next one. Imagine the volts (which produce the sound) jumping up and down, which is nowhere near the sine you drew. But it it not about that ... What is, is this :

You assumed the sampler starting at the starting point of the sine. But why would it ? So, assuming your drawing is 10cm or 4 inches wide, create your first sample point 1 cm to the right of the original one. Shift the other samples 1 cm to the right similarly. Now, do you see what happens ? you have the same base frequency, but the voltages (on the Y-axis) have become totally different. And you bet the tone coming from that is totally different as well. And the only thing what actually happened, is after the A/D recorder started recording, the pianoplayer started 1us later playing opposed to your first "recording" (sample drawing). Now, if the piano player would be able to play very consistently (hence the time difference with the sampler wouldn't change all over the track), the sound will be TOTALLY different. Same recording engineer, same piano player, same recording device, same everything but on thing not under control : time shift compared to the start of the sampler.

In either case it is the intention of it all that your nicely drawn sine is reconstructed. You know, that sine which already isn't in the CD data anymore ...

Never minding Arc Prediction, now think again about the bunch of next samples contributing to the current one to be output like normal filtering does. Now think of not the pianoplayer starting to play at a fraction of a ms later, not the sampler starting later etc., but your EAC rip not starting at the offset as "intended". Let's say the offset is one sample later only. Look at your sine again. When it starts at the right place the first value to calculate with is 0. But one sample later it is 37767 (as the digital representation of 2V redbook). Ouch. What will that do to the calculation of the next output samples, THROUGHOUT the track ?? it will all sound different !

Things may be not as black and white as I suggest, but they all influence ...

Peter

[AUDIODIDAKT]

FYI,Did you know there is software to restore those drive offsets in flac files ? (to AccurateRip compliant)
Maybe do some testing
It can also recognize HDCD's and many more, love it !

Its called Cuetools 2.0.6

[manisandher]

Also, if the Microsonics would have worked with pure multibit A/D's I really wonder which that ever would have been for 192KHz (up to 2002). Did 24/192 multibit chips ever exist ?

Hi Peter,

Here's what Michael Ritter (of Pacific Microsonics) wrote in 'Mix' magazine in 1999, before the Model Two was released (the highlights in bold are mine):

"The actual A/D converter in the Model One runs at 24 bits and 176.4 kHz currently; the Model Two will also convert at 192 kHz. We improve the linearity of our conversion with a high-amplitude broadband dither signal that we mix in with the program in the analog domain. The dither appears to be random, but the system knows at any given instant precisely what the amplitude of that dither signal is. And because we use our own custom, discrete, full-ladder converter with excellent amplitude and phase accuracy, we are able to apply an 'anti-dither' signal, exactly out-of-phase and matched in time, in the digital domain after conversion. That nulls the dither noise out of the signal.

If it's going to be a 176.4 or 192kHz DVD-Audio release, then we will not decimate that signal; we use a proprietary filter [non-oversampled] optimized to that sample rate. If it's going to be 88.2/96 kHz, we use 2:1 decimation, and once again we use a filter optimized to that frequency. But in both high-resolution settings, the Nyquist frequency is high enough that we don't use the 'dynamic decimation' process that becomes necessary when we go down to 44.1 or 48 kHz."

Note that there may be some confusion with regard to the editing by Mix magazine. Ritter mentions a 'proprietary filter' and the editor has included '[non-oversampled]'. I'm not sure how accurate this is...

HTH.

Mani.

[manisandher]

Ok, turning this some other way around, at listening to random (non-HDCD) HRx recordings, they are not a bit better, they are HUGELY better.

Yes, my experience too. But I think you need a 24/176.4 NOS DAC to appreciate this...

This weekend, I'll try to find some time to make some recordings from vinyl for you - 16/44.1, 16/44.1 with peak extension, 24/176.4 and 24/192. You won't like the music much (it'll be classical or jazz - that's all I've got on vinyl), but hopefully it might help you understand what's going on a bit better - or at least eliminate Keith Johnson's recording techniques from the equation.

Cheers,
Mani.

[PeterSt]

But key is the "full ladder" thing (which I missed thus far in my readings). So, I will believe that, but which chip was (at that time) able to do it. So, this goes back to before 1999, although possibly 176.4 only for that chip. The latter makes it kind of suspectible, because (at that time !) nobody was dealing with 88.2 or 176.4. Of course, the useable derival of it, 44.1, was.

Hmm ... Did you actually ever try to open the box and look for chips ? (I am not saying you should do this now on behalf of this little discussion !). Or, if it's easy to open the box you may take a photo of it and send it to me (maybe by normal email because of who knows what's an actual secret in there).
But again, you may have better things to do !

Thanks,
Peter


PS: About your recordings ... mwah, that might lead to something. But I don't know yet. You may leave 192 just as well out.

PPS: It intrigues me. I just won't believe Keith stopped all this because of parts running obsolete. Just as that I won't believe it will take a million to restart production with the currently available parts. Well, let's say that I didn't consume a million yet. Haha.