Ja, jag är en smula "upphängd" på Chord Electronics just nu, och konstruktören Rob Watts i synnerhet, därav ett antal återkommande inlägg i ämnet. Behöver hjälp med att granska några av hans utsagor och teorier, där min egen kunskap inte räcker till längre. Kan redan nu be om ursäkt för att jag med största sannolikhet kommer använda mig av nomenklatur som i viss mån kommer från marknadsföringen.
Vi har haft intressanta och väldigt givande trådar med granskande av MQA och lite om DAC-teknologier, så hoppas denna skulle kunna bli något liknande. Ska försöka samla ihop det jag undrar över på ett så bra vis som möjligt. Det blir dock mycket text, då jag tycker man kan behöva se mer än brottstycken. Inläggen nedan handlar om olika saker, ibland är ämnet DAC:ar, ibland ADC:er, men har hans idéer om de problem hans teknik ska lösa gemensamt.
Rob Watts verkar förespråka blind AB-lyssning när han utvärderar sina konstruktioner, vilket känns sunt, men han verkar också försvinna iväg på subjektivt gungfly då och då samt egenkonstruerade teorier om hur hjärnan processar ljud osv. Det jag undrar om det kan finnas någon sanning i, är hur/om hans avancerade noise shapers, rekonstruktionsfilter skulle kunna återge något som tidigare inte kunnat återges. Det han mestadels menar att hans FPGA-teknologi med noise shapers och filter lyckas med att återskapa är originalets transienter samt det djup som finns gömt i inspelningen. Och lyckas göra detta bättre än vad annan existerande teknologi lyckas med. Om det ens är teoretiskt möjligt är vad jag skulle vilja resonera om. För mig låter det plausibelt att det skulle kunna vara möjligt med en teoretisk förbättring, men har rekonstruktionsfilter innan varit så dåliga att denna teoretiska förbättring även drastiskt skulle höras?
Det är ett intressant spörsmål tycker jag, om man leker med tanken att "japp, det stämmer, förbättringen hörs". Det skulle (väl?) innebära att om vi tar exemplet med LTS lyssningstester, att en jämförelse mellan referens D/A-omvandlingen och Chords D/A-omvandling skulle uppfatta en skillnad, men där skillnaden i så fall inte skulle innebära "färgning" utan tvärtom. Är det ens en rimlig tankegång? Slå gärna hål på den.
Några produkter nämnas i inläggen nedan, kanske lika bra att reda ut vilka de är:
Dave: Deras top-of-the-line DAC
Davina: Kommande ADC.
Hugo: Instegs-DAC
Rob Watts: Frustrated by the performance of DAC chips, in 1994 I acquired digital design skills and started designing my own DAC's creating pulse array DAC technology using a new device called FPGA's. These devices allowed one to make your own digital design by programming an FPGA. I then became an independent design consultant, and started working with Chord, and the first product was the DAC 64. This was unique, in that it was the first long tap length FIR filter (the WTA filter).
[…]
The invention of the WTA filter with the DAC 64 nicely exposes the conventional engineering assumption - that the length of an FIR filter (an FIR filter is used in DAC's to convert the sampled data back into a continuous signal) does not matter, that all we need is a suitable frequency response. But if one looks into the maths of sampling theory, then it is clear that to perfectly recover the bandwidth limited signal in the ADC then an infinite tap length FIR filter is needed. It is also obvious to me that if you had a small tap length filter, then the errors would present themselves as an error in the timing of transients. Now a transient is when the signal suddenly changes, and from my physiology of hearing studies transients are a vital perception cue, being involved in lateral sound-stage positioning, timbre, pitch perception and clearly with the perception of the starting and stopping of notes.
Från tråden Watts Up...? @ Head-Fi. Intressant inlägg i sin helhet.
Rob Watts: Going back to sound stage - no there are no measurements for sound stage directly (excepting gross errors). Now I associate depth with small signal non-linearity - but my DAC's have had perfect small signal linearity for many years. Indeed, before Dave, I thought there was no more improvements to be had with depth, in that the losses we hear with reproduced sound is down to the ADC and mastering. That was not an unreasonable assumption, as Hugo noise shapers already had 1000 times more small signal accuracy than conventional noise shapers, and zero measurable fundamental linearity (to do this I had long ago given up on traditional measurements - I use a -120dB signal and measure how accurate it is at -120dB using an FFT).
But the thing about assumptions is one needs to test them with listening tests - and with Dave I had loads of FPGA space, so could have better noise shapers - and increasing the performance of the noise shaper did improve depth. But it changed no other aspect of sound quality. In the end, I had noise shapers that were a trillion times better than conventional, and at each improvement in noise shaper performance it largely (sometimes it didn't but that's way too complex to talk about) had deeper depth.
But although we are limited with measurements, with digital design work you can run Verilog simulations which will give you bit perfect data on the performance of your digital circuit - so you can look at very small errors, then listen to them via the FPGA. So we can actually quantify (measure) very small errors, then see if they have any audible consequences. If you can put a number on an effect, then it means you really understand the phenomena, and modern tools allows this kind of analysis.
Från tråden Watts Up...? @ Head-Fi
Rob Watts: A perfect sinc response infinite FIR filter will perfectly reconstruct the original bandwidth limited analogue signal - without any change whatsoever - no pre ringing or post ringing, it is mathematically identical except displaced in time. But such a filter has an infinite amount of pre and post ringing, so looks nothing like the original impulse?
How can you explain that paradox that a filter that won't change the original signal at all has an infinite amount of ringing? This is actually a very important question as the whole audio industry in trying to minimize ringing has got the filter question completely and utterly wrong.
Från tråden Watts Up...? @ Head-Fi
Rob Watts: I suppose the point I was getting across that for a DAC interpolation or reconstruction filter, I have never seen the ringing performance being a significant factor; by this I mean there were always other explanations for the differences in SQ I can hear whilst optimizing the SQ during listening tests on the WTA algorithm. In broad terms, getting the filter to be close to the ideal sinc response is what you are doing - but when optimizing the sound quality sometimes that will mean more ringing, sometimes less ringing - its just not a relevant factor. The reason for this is that an ideal filter will ring infinitely, but will recover the un-sampled bandwidth limited signal absolutely perfectly with no change to the signals ringing at all. If we want to evaluate a filters performance using ringing, then we need to use a bandwidth limited impulse response test signal. So the ringing performance using an illegal (from sampling theory) impulse response has no bearing on the SQ - except that the more ringing, then perhaps the closer it becomes to the ideal sinc filter. Certainly going for a filter that does not ring is a very bad idea - the ringing is absolutely essential as it is the multiple rings from all the past and future samples that when added up gives the correct transient values.
Just to give you an idea - I have designed a 8FS WTA filter, and a 16FS WTA filter with identical ringing performance - except the 16FS filter has of course double the output resolution - and when you see the two impulse responses after they have been further filtered to 2048 FS they look identical - but the two filters actually sound very different.
So the filter for the interpolation filter has to be a certain way with lots of ringing, if you want to recover the signal perfectly. But for the ADC the only requirement is that it is perfectly bandwidth limited, theory does not care how it is done. So I plan to just try different filters to define the aliasing problem (which I know for certain from listening tests is a major SQ issue) and then try differing ringing responses to see which type of filter sounds best. My current thinking is that ringing in an ADC will turn out to be unimportant; we just need it to be bandwidth limited by enough to eliminate aliasing, but I too am worried about the effect of 22 kHz bandwidth limiting for redbook.
Från tråden Watts Up...? @ Head-Fi
Rob Watts: Generally I have two primary motivations - getting closer to the un-amplified sound I hear in concert halls - or better sound quality is one. The second is about improving my understanding and knowledge. Davina mostly in the short term is about better understanding - it could take some time before we have substantial numbers of new recordings available. On the understanding side I have a number of issues I want answers too:
1. How much aliasing is acceptable? Conventional ADC's often use half band filters, and these have aliasing that peaks at -6dB. The argument is that you can't hear anything from 20 to 20k, so it does not matter. The fact that the aliasing has completely destroyed the timing of transients has not occurred to them. In my experiments, previous decimation had -140 dB performance, and I know that is not good enough measurement and SQ wise.
2. Will we get lifelike depth reproduction? I want to hear an organ at 100m away sounding like its 100m away in your living room. For sure, a problem is noise shaper resolution, and I have designed 350dB (Dave standard) analogue noise shapers for Davina. Is this the whole answer?
3. Can we decimate then interpolate with WTA completely transparently? This will give an absolute indicator of how good the WTA filter actually is.
4. How do we get 16 bit data but with no depth problems? Will triangular dither do it perfectly?
Technically, I have already solved the issues of noise floor modulation, which is a major problem with ADC's - and I can do this because the analogue is all discrete. Also, distortion is mostly about the reference DAC, and I know from Dave how to solve that problem too. But there remains a number of very interesting issues to understand.
Från tråden Watts Up...? @ Head-Fi
Rob Watts: Let's take a step back, and examine the scale of the problem. In order to solve the resolution problem, what do we need for perfect transparency? And what level do we need for perfect timing?
Ask that question, how much resolution do we actually need will depend on doing listening tests to find out - theory won't help you as we have no idea how the brain process's ear data to determine depth for example. You can only do it by designing noise shapers and actually listening to them. In the Dave project I did exactly that and was shocked to the core by what I discovered - there is no limit to how good the noise shaper needs to be.
That's actually quite a profound statement. I can substantiate this by saying that a 200 dB (THD and noise 20 kHz BW) has much better depth perception than a 220dB one. Furthermore, listening to a noise shaper with better than 350 dB has audibly deeper depth than a 330 dB one (albeit it's getting to be a much smaller change). These are staggeringly small numbers - better than 350 dB has a trillion times more resolving power than DSD 64 - no wonder DSD has such poor depth perception - the noise shaping is simply not good enough at resolving small signals. Indeed, the required performance is still 30 billion times better than the PWM example I talked about earlier. Now it's possible that the requirement for better than 350 dB performance on the noise shaper was simply a proxy for something else going on in the analogue section - but that when I use this performance for digital noise shapers that are purely digital (truncation noise shaping) I still hear the benefits in terms of depth reproduction. So the evidence based on my listening is that the brain is sensitive to small signal amplitude error of any size - even if almost zero and we need > 350 dB noise shaping
What can we say about timing errors?
If you would have asked me this a few years ago, I would have said uS accuracy was needed. Now I make no such assumption - there is perhaps no limit to how good the timing of transients need to be. So how can I substantiate that bold statement? Unlike noise shapers, it's rather difficult to put a number to timing accuracy. I guess I ought to state what I mean by transient timing accuracy. I do not mean - unlike the rest of the audio business - ringing performance; this is absolutely not what I am thinking about when I talk about the time domain or timing accuracy. Ringing uses an illegal signal from sampling theory POV as it is not bandwidth limited, so you would not actually get a perfect impulse from a perfect legal bandwidth limited ADC. So why worry about a signal you will never get? So it is actually pointless talking about it. What I mean is the accuracy of the timing of transients. Imagine a bandwidth limited analogue signal that is being sampled in the ADC - it is fully negative, goes positive and at some time crosses through zero. Let us say it is sampled at 44.1 kHz, so every 22,676 nS it's sampled. Let us imagine that the signal is sampled, and then crosses through zero at exactly 20,155 nS after sampling. Of course, when it gets sampled again at +22,676 nS it will now be a positive value. The question is, when the DAC reconstructs the sampled data - converting sampled data back to a continuous analogue signal - when will the signal cross thru zero? Theory is completely clear and undeniable - if we use an infinite oversampling FIR filter with a sinc response at 22,676 nS and a perfect DAC we will reconstruct the time it crosses thru zero absolutely perfectly at 20,155 nS. But with a finite non sinc function reconstruction filter, it will not cross thru at exactly 20,155 - maybe at 19,000 nS or 21,000 nS. And it is these differences in the timing of transients, are what I am talking about. Now in the past I would have said that getting it right to a uS was perhaps OK (timing errors can be as big as 100uS in conventional filters) - now I know that instead of worrying about uS we need to worry about getting it correct to nS's.
Från tråden Watts Up...? @ Head-Fi. Intressant inlägg i sin helhet.
Rob Watts: The timing I am talking about is the reconstruction of transients. You can see this by imagining a transient in an ADC. It is -1v at the sample point, and +0.5v at the next sample point. It crosses through zero at say 15,679 nS after the first sample point. Now assuming the signal in the ADC is perfectly bandwidth limited, then a infinite tap length sinc function filter will exactly reproduce it crossing through zero at +15,679 nS. But with a conventional DAC it will not cross through at +15,679 nS - but could vary by many thousands of nS - and this variation or inaccuracy is highly audible, with very important musicality consequences. It is this timing error within interpolation filters (the filter that takes it from 44.1k up to 705.6k and eventually to 104 MHz within my DAC's) that I am talking about.
Från tråden Watts Up...? @ Head-Fi
----------
Inlägget är redan på tok för långt, så jag stannar här. Ska se om jag kan hitta några andra intressanta inlägg senare. Chord har precis släppt nyheten om deras CD-spelare Blu Mk 2 som de säger, så klart, ska ge drastiska förbättringar på red book-material. Kan det ens vara möjligt att gräva fram mer eller har vi hört allt på våra CD-skivor?
