[Jawed]

As I understand it. Mojo 2 performs EQ once WTA1 has been applied, i.e. with music data at 705.6khz or 768khz, these being 16FS (16x the base frequency of either 44.1khz or 48khz).

We should expect the new scaler to output 16FS, in order to be compatible with Mojo, Mojo 2, Qutest, Hugo 2, Hugo TT2 and DAVE.

So, in theory, there's an opportunity to apply EQ on the output of the new scaler in the same fashion as seen in Mojo 2.

In my opinion this creates a problem though, in which some taps or other processing was sacrificed for EQ. Now it might be that the hardware cost of the EQ is so low that it would be equivalent to only 5 to 10 extra taps, say. Fairly inconsequential in the context of 1 million, or much more, taps. I have no idea.

The same question applies to Mojo 2 but it was deemed so important it was added, even if sacrificing some taps...

 

[Currawong]

As for successors to the Dave, there is also the issue of the UI that would be used to control the eq function. Would this be via ever more complex external hardware and displays, or the development minefield of a dedicated app?

For Chord there is again a real danger of severe mission creep. A repeat of the GoFigure experience is not needed this time around.

Don't forget that the Poly and 2go are not actually made by Chord, but a third party, and Rob is not involved with either of them. So, I highly doubt that any external software would be involved. Rather if, say a DAVE with EQ were made, I imagine it'd be rather similar to the Mojo 2 in how it works and what it can do, but with the better controls and screen of the DAVE.

I also imagine that since Rob already has the basic building blocks for any DAC of his already there, he doesn't have to redesign the code so much as decide how much, what and where on the FPGAs, design the hardware for it, and do many listening and other tests. I'm of the impression that he's continuously refining the WTA filter as well, so there's also that.

 

[Rob Watts]

I'm not sure I see the use case for a full parametric eq implemented in a desktop or full size DAC.

For a portable system ok as the eq function needs to go out and about with the rest of the system.

For non-portable systems wouldn't it be better to implement your eq at source via eg convolution filters or the equalisers provided in the streaming apps, which all have the control software and apps necessary already developed?

As for successors to the Dave, there is also the issue of the UI that would be used to control the eq function. Would this be via ever more complex external hardware and displays, or the development minefield of a dedicated app?

For Chord there is again a real danger of severe mission creep. A repeat of the GoFigure experience is not needed this time around.

I say: keep it simple. A DAC is a DAC. Eq should be implemented elsewhere.

Full parametric EQ wouldn't be a good idea - intellectually I don't like them, as using resonators to cancel other resonances is impossible to do properly. Moreover, the vast majority of Chord users would be incapable of using it properly. It's always better to solve the issue at source than apply band-aids.

But having adjustable EQ that does not destroy SQ or more importantly musicality I think is absolutely vital. And I find that standard EQ is incapable of sounding transparent - you are better of not using it. And don't get me started about the SQ problems of convolution!

As to mission creep, that is a danger, but that would not come from Chord as they don't tell me what to do on the insides. So it was my idea to spec Mojo 2 EQ and design the 104 bit EQ core that's inside Mojo 2. I am not good at estimating how long things take to do, and I am a perfectionist, so will only release stuff when I am perfectly happy with SQ. I remember spending an extra month on an issue, and heard improvements, but there would be no way the end user would notice it.

And I agree - keep it simple (KISS) is my byword - but performance always trumps KISS.

As I understand it. Mojo 2 performs EQ once WTA1 has been applied, i.e. with music data at 705.6khz or 768khz, these being 16FS (16x the base frequency of either 44.1khz or 48khz).

We should expect the new scaler to output 16FS, in order to be compatible with Mojo, Mojo 2, Qutest, Hugo 2, Hugo TT2 and DAVE.

So, in theory, there's an opportunity to apply EQ on the output of the new scaler in the same fashion as seen in Mojo 2.

In my opinion this creates a problem though, in which some taps or other processing was sacrificed for EQ. Now it might be that the hardware cost of the EQ is so low that it would be equivalent to only 5 to 10 extra taps, say. Fairly inconsequential in the context of 1 million, or much more, taps. I have no idea.

The same question applies to Mojo 2 but it was deemed so important it was added, even if sacrificing some taps...

It's actually more complex and finely balanced than that. So my custom EQ DSP core is not insignificant in gate count. But it only uses 4 multipliers plus some FPGA fabric, and relatively small amounts of memory. But on the other hand, it saves on multipliers and FPGA fabric, as the DSP core for cross-feed and volume control, SPDIF EQ is no longer needed. Given that with the substantial sound quality benefits of transparent EQ it becomes a no brainer to employ it.

I also imagine that since Rob already has the basic building blocks for any DAC of his already there, he doesn't have to redesign the code so much as decide how much, what and where on the FPGAs, design the hardware for it, and do many listening and other tests. I'm of the impression that he's continuously refining the WTA filter as well, so there's also that.

Absolutely. I never imagined twenty years ago I would still be learning more and improving SQ with DACs now!

 

[Kentajalli]

Full parametric EQ wouldn't be a good idea - intellectually I don't like them, as using resonators to cancel other resonances is impossible to do properly. Moreover, the vast majority of Chord users would be incapable of using it properly. It's always better to solve the issue at source than apply band-aids.
But having adjustable EQ that does not destroy SQ or more importantly musicality I think is absolutely vital. And I find that standard EQ is incapable of sounding transparent - you are better of not using it. And don't get me started about the SQ problems of convolution!

Have you ever considered making a multi-room server/player?
And then, something like a TT3 with network capabilities, possibly a simple EQ DSP, for each room! who knows, perhaps with a 50W class D inside! Just about enough for a bedroom. If one needs more, let'em buy an external amp.
I think such a TT3 would sell like hotcakes, if priced right (spelt LOW enough ).
A one box solution, with EQ, amp and network DAC, no more USB or electrical SPDIF noise, sitting neatly on a table.
Others have tried, but if you set your mind to it, that would be end-game.

 

[ChrisGB]

Personally, I prefer the approach of just making a DAC. Going Swiss Army knife never gets you as good a blade as a dedicated knife!

The other question that this post prompted my brain to ask is this:

If we hear sonic differences in processes that are managed in the 768KHz region, what does a class D amplifier working in potentially similar frequency areas do to the timing of transients?

 

[Kentajalli]

Personally, I prefer the approach of just making a DAC. Going Swiss Army knife never gets you as good a blade as a dedicated knife!

One of the attractivenesses (let's call it that) of a TT2, is its ability to drive efficient speakers out of box.
By implementing networking, one is isolating the DAC completely, yet through a single interface you can have your 24/768kHz and eat it.

If we hear sonic differences in processes that are managed in the 768KHz region, what does a class D amplifier working in potentially similar frequency areas do to the timing of transients?

You are assuming, we hear such differences, we may not. RW being RW, he can perhaps drive the output devices, direct from the digital output, making it a digital amplifier, that is if he would be so inclined.
With a digital amp, or class D output stage, a possible TT3, can be small in dimensions, light weight, and produce very little heat!
All in a manageable, small enclosure. That would take minimalist audio systems, to new levels. Other manufacturers do have such devices, but as yet none can compete with Chord sound.
Imagine if such a TT3 could exist! I buy that for a dollar(well, a few dollars).

 

[Rob Watts]

Have you ever considered making a multi-room server/player?
And then, something like a TT3 with network capabilities, possibly a simple EQ DSP, for each room! who knows, perhaps with a 50W class D inside! Just about enough for a bedroom. If one needs more, let'em buy an external amp.
I think such a TT3 would sell like hotcakes, if priced right (spelt LOW enough ).
A one box solution, with EQ, amp and network DAC, no more USB or electrical SPDIF noise, sitting neatly on a table.
Others have tried, but if you set your mind to it, that would be end-game.

No - my only focus is on improving sound quality, and to close the huge gap from acoustic live sounds to reproduced sounds. Multi-room server/player won't help with that goal - I will leave that to others!

Personally, I prefer the approach of just making a DAC. Going Swiss Army knife never gets you as good a blade as a dedicated knife!

The other question that this post prompted my brain to ask is this:

If we hear sonic differences in processes that are managed in the 768KHz region, what does a class D amplifier working in potentially similar frequency areas do to the timing of transients?

If you need EQ - and I use it with my Stealths - then the issue is to do it transparently, and that means doing it at 16FS, employing noise shaping at each node, and using large fixed point bit depths and completely avoid floating point. The only practical way of doing all that is via custom designed hardware...

And absolutely Class D suffers from transient timing distortions, that's why it sounds poor. The sensitivity of the ear/brain to transient timing distortions is very high and you need to go very much higher in frequency than just 16FS (or 705/78kHz) to deal with it. You would be amazed at how minute errors at 2048FS can make to the sound quality.

One of the attractivenesses (let's call it that) of a TT2, is its ability to drive efficient speakers out of box.
By implementing networking, one is isolating the DAC completely, yet through a single interface you can have your 24/768kHz and eat it.

You are assuming, we hear such differences, we may not. RW being RW, he can perhaps drive the output devices, direct from the digital output, making it a digital amplifier, that is if he would be so inclined.
With a digital amp, or class D output stage, a possible TT3, can be small in dimensions, light weight, and produce very little heat!
All in a manageable, small enclosure. That would take minimalist audio systems, to new levels. Other manufacturers do have such devices, but as yet none can compete with Chord sound.
Imagine if such a TT3 could exist! I buy that for a dollar(well, a few dollars).

I would never use Class D - it's not fit for purpose to recreate the sense of real musical instruments playing with a sense of space. But higher powers from a single amplification DAC is absolutely possible - and if one wished, using Class H, capable of high efficiency too.

 

[Reactcore]

If you need EQ - and I use it with my Stealths

I thought the Stealth was already as ideal (flat) tuned as it gets..

 

[rkt31]

My question to rob. Is there any advantage of going for 256 bit processing like done by pggb latest offline version? They claim to solve the limitiation of windowing.

 

[Kentajalli]

No - my only focus is on improving sound quality, and to close the huge gap from acoustic live sounds to reproduced sounds. Multi-room server/player won't help with that goal - I will leave that to others!

Thank you for replying.
I think you already have improved sound quality by a far margin.
I assumed, doing away with electrical data connection, such as USB and SPDIF would solve some possible noise issues, not to mention interconnects. Also, since TT2 can drive efficient speakers well, then having such a device with a bit more real power, no electrical noise (or white noise blast!) would go towards improving sound quality.
Sound quality is a loose word, because it relies on the entire chain, speakers and amplifiers included. I reckon Chord is not going to speaker manufacturing, but amplifiers, they already do. TT2 has a token amplifier built in ( OK high current output section), why not go a little further.
Let's forget about servers! how about just network capability?
I am not trying to push you, just gently nudging .

I would never use Class D - it's not fit for purpose to recreate the sense of real musical instruments playing with a sense of space. But higher powers from a single amplification DAC is absolutely possible - and if one wished, using Class H, capable of high efficiency too.

OK, no class D, but class H? isn't that more expensive, more complex?
How about just 50-75W class AB then? that's enough for most bedrooms, or average small UK living rooms.
A one box solution, network capable, is very attractive to me, and possibly a lot of minimalist Hifi consumers, who just want to listen to music, and not interested in boxes and cables and PSU's and . . .

 

[ecwl]

My question to rob. Is there any advantage of going for 256 bit processing like done by pggb latest offline version? They claim to solve the limitiation of windowing.

When PGGB first came out and I was demoing it, I told the programmer of PGGB that I found the apodising filter for 44kHz files to worsen transient responses so he allowed me to use the non-apodising filter that fixed the problem but he told me none of the beta testers found it to be a problem and they all preferred the apodising filter.

I also told the programmer that I found the low level detail and clarity and soundstage depth to be lacking compared to M-Scaler. People online basically told me the problem was the USB source or Chord’s USB chip which I thought sounded a bit insane. Since I was using the same source for M-Scaler or DAVE or Hugo 2 or Mojo for testing PGGB. Of course I wonder if these beta testers can hear what I heard. I asked the programmer the following question: is it possible that floating point FFT convolution at 64-bit has worse accuracy than direct convolution at 64-bit. I never got an answer except I was told to go out and buy a $1000 USB to dual BNC adaptor. And to pay for PGGB afterwards. Of course now, PGGB does convolution at high bit depth. Makes me wonder if that’s the reason. But my math on FFT floating point convolution isn’t good enough to know if I was correct.

 

[Progisus]

When PGGB first came out and I was demoing it, I told the programmer of PGGB that I found the apodising filter for 44kHz files to worsen transient responses so he allowed me to use the non-apodising filter that fixed the problem but he told me none of the beta testers found it to be a problem and they all preferred the apodising filter.

I also told the programmer that I found the low level detail and clarity and soundstage depth to be lacking compared to M-Scaler. People online basically told me the problem was the USB source or Chord’s USB chip which I thought sounded a bit insane. Since I was using the same source for M-Scaler or DAVE or Hugo 2 or Mojo for testing PGGB. Of course I wonder if these beta testers can hear what I heard. I asked the programmer the following question: is it possible that floating point FFT convolution at 64-bit has worse accuracy than direct convolution at 64-bit. I never got an answer except I was told to go out and buy a $1000 USB to dual BNC adaptor. And to pay for PGGB afterwards. Of course now, PGGB does convolution at high bit depth. Makes me wonder if that’s the reason. But my math on FFT floating point convolution isn’t good enough to know if I was correct.

Pggb to my ears adds air which is a pleasant effect. I do not hear this when using the mscaler. I have no idea if this is accurate or not but it is pleasant. I listen to my local files with pggb 256b and streamed with the mscaler. The best of both worlds. It does require two streamers though. One to the mscaler and dual bnc to tt2. The other usb to the tt2.

 

[Jawed]

I asked the programmer the following question: is it possible that floating point FFT convolution at 64-bit has worse accuracy than direct convolution at 64-bit.

At the same precision of the mathematical operations, the FFT route will be more accurate simply because the count of operations is lower, assuming that there's a long tap length. Only for shorter tap lengths would direct convolution be more accurate, as the number of operations would be lower.

https://ccrma.stanford.edu/~jos/mdft/Convolution_Theorem.html

[EDIT: added this section] This is also a good post on this topic, when considered for "real time" filtering:

https://dsp.stackexchange.com/a/71280

 

[GuiltyRocker]

No - my only focus is on improving sound quality, and to close the huge gap from acoustic live sounds to reproduced sounds. Multi-room server/player won't help with that goal - I will leave that to others!



If you need EQ - and I use it with my Stealths - then the issue is to do it transparently, and that means doing it at 16FS, employing noise shaping at each node, and using large fixed point bit depths and completely avoid floating point. The only practical way of doing all that is via custom designed hardware...

And absolutely Class D suffers from transient timing distortions, that's why it sounds poor. The sensitivity of the ear/brain to transient timing distortions is very high and you need to go very much higher in frequency than just 16FS (or 705/78kHz) to deal with it. You would be amazed at how minute errors at 2048FS can make to the sound quality.



I would never use Class D - it's not fit for purpose to recreate the sense of real musical instruments playing with a sense of space. But higher powers from a single amplification DAC is absolutely possible - and if one wished, using Class H, capable of high efficiency too.

Rob just confirmed why I have never liked class D, it has never sounded natural to me, now I know the reason. Thanks Rob

 

[ecwl]

At the same precision of the mathematical operations, the FFT route will be more accurate simply because the count of operations is lower, assuming that there's a long tap length. Only for shorter tap lengths would direct convolution be more accurate, as the number of operations would be lower.

https://ccrma.stanford.edu/~jos/mdft/Convolution_Theorem.html

[EDIT: added this section] This is also a good post on this topic, when considered for "real time" filtering:

https://dsp.stackexchange.com/a/71280

Interesting. I saw what you linked to. Can you explain further? Because what I saw is that FFT convolution is faster than direct convolution. And that FFT convolution has fewer operations (addition and multiplications) than direct convolution. And yes, I guess with enough additions and multiplications, you’re going to get less accurate because of rounding errors with each addition and multiplication operation. So I get that part. But of course, maybe that can be mitigated, not sure.

But what about the actual FFT and reverse FFT themselves? Do they contribute to errors and bit depth inaccuracies? Can you provide the math for that? Or maybe a link to that?