Audio Asylum Thread Printer
Get a view of an entire thread on one page
|For Sale Ads|
In Reply to: RE: To clarify ... posted by Jim Austin on June 05, 2017 at 13:06:30
>> There are two spectra in the chart below: MQA is red (although you can hardly see it) <<
I can only see the yellow line and therefore assume that it overlays the red line almost exactly. If that is the case, then something is very odd with the digital filter in the D/A converter. There is no way that a "leaky" digital filter with a slow rolloff would create that spectrum.
The only possibility I currently see is that you listened to "the CD-res file, with the MQA decoder turned on--but this of course is not an MQA file, so that shouldn't matter". Apparently it does matter, which would imply some things that don't seem at all right to me. Before going down that rabbit hole, could you do us a favor and measure the spectrum of the CD with the MQA filter disengaged?
PS - I also had a similar concern when (the other) JA measured the Mytek Brooklyn. Figure 6 (linked below) shows the response of the MQA filter. I do not understand how it would be possible to measure the MQA filter without having a special MQA test tone that would activate the MQA filter. Again, more questions than answers.
I'll do this when time allows. I should make it clear though, having measured many more MQA files and compared with non-MQA versions--that this is not typical for MQA, although it may be true that 44.1 MQA files do closely resemble their CD equivalents; so far I've only measured one other: first track from Sonny Terry & Brownee McGee's Live at the New Penelope Café. MQA is in red.
Thanks for posting this second file. It provides some good clues to help understand what is going on. There are still a lot of variables remaining, so it is difficult to draw any conclusions at this point. However we can say that:
1) Apparently it is possible to create an MQA file of a 44.1 kHz track that doesn't have a marked notch centered at Fs/2. Currently it is unclear if this is due to the settings during the MQA encoding or possibly some differences in settings on the playback DAC you used.
2) They cyan trace is presumably that of the CD. I have *never* seen the spectrum from any CD that did not have a sharp drop-off at Fs/2 due to the anti-aliasing filter in the A/D converter. This trace does not. Even if there were *no* filter in the playback D/A converter (ie, a "non-oversampling" design), the spectrum should *not* look like this at all. What D/A converter did you use to capture this spectrum, and how was it configured?
> > What D/A converter did you use to capture this spectrum, and how was
it configured? < <
Mytek Brooklyn. Until I reviewed (the other) JA's measurements, which you posted a link to, I had forgotten that you could choose your own filter: Fast, slow, MPh. These choices apply only when the MQA filter is disabled. So I checked. It's set to MPh. MQA filter is also MPh. I guess it's the same MPh, or very close.
> > you could choose your own filter: Fast, slow, MPh. These choices apply only when the MQA filter is disabled. So I checked. It's set to MPh. MQA filter is also MPh. I guess it's the same MPh, or very close. < <
I still don't fully understand what is happening here:
1) The Mytek Brooklyn uses the ESS ES9018K2M DAC chip. I am very familiar with it as it is the same model that was very first used in the Pono Player. It has a choice of three stock digital filters built into it, which correspond to the choices offered by the Brooklyn:
The top graph is the "Fast" filter, the middle graph is the "Minimum Phase" filter, and the bottom is the "Slow" filter. The ESS chip is somewhat unusual in that its internal filter is a concatenation of a 4x initial stage followed by a 2x stage for a total of 8x. (Virtually all other DAC chips use a concatenation of three 2x stages.) As you can see, the frequency response of the Minimum Phase is virtually identical to the Fast filter. Instead of having ±10 cycles of both pre- and post-ringing, it has ~20 cycles of post-ringing only (there's no free lunch).
NB: You can estimate the number of taps in each stage by counting the number of "bumps" or "ripples" in the stop-band response. The "Slow" filter has about 15 taps and it appears that all calculations are done in one pass. The first (4x) stage of the Fast and Minimum Phase filters probably has almost 100 taps, while the second (2x) stage appears to have about 9 taps. This is a total of (say) 109 taps, whereas doing it in a single pass with the same frequency response of the first section would require 200 taps - which is why companies concatenate filters - it saves money. (These are just some of the "fingerprints" left behind by digital filters that give us clues into what is causing the spectra you have been posting.)
But the MQA filter looks nothing at all like any of the ones built into the ESS DAC chip. That filter's behavior is clearly seen in JA's review, linked earlier. Furthermore my understanding of the MQA digital filter is that it is implemented in an external DSP chip, typically an XMOS - which is what is found in the Mytek Brooklyn.
It still seems something is amiss. I am completely confused that (the other) JA was able to measure the response of the MQA filter, apparently without access to a special MQA test file that would activate the MQA filter. It also seems that somehow the MQA filter was active in the second spectrum you posted. Normally all CD recordings will yield a spectrum like the first one you posted, where the energy falls off a cliff right around 20kHz. If MQA were more forthcoming in the details of how their system worked, we wouldn't have to scratch our heads on this. If you have the time and interest, I would suggest a couple of further tests you could try:
1) Run the spectrum of the CD on the Brooklyn with the MQA both engaged and disengaged.
2) If you don't see a difference there, try comparing the CD spectrum from the Brooklyn with the spectrum obtained from a different D/A converter.
3) If you want to get really hard core, have (the other) JA send you a copy of the -4dBFS white noise file he uses to recreate the test Juergen Reich of MBL created. This allows one to examine the stop-band response of a digital filter, allowing you to remove the variable of the spectral content of the music on your CD.
As always, solely my personal opinions and not necessarily those of my employer or your employer.
No new measurements, just new analysis.
If you look closely at the FFT I posted, you'll see a small double peak near 25kHz. I went back to Audition so that I could get a closer look and found that the highest of those two peaks is at 24.75kHz, at a level of -89.76. So do the math: That's Nyquist +2.7kHz.
So what's happening at Nyquist - 1.7 kHz? Notice anything interesting? Frequency resolution is limited, even in Audition, but there's obviously a very similar-looking double peak at 19.35kHz, or Nyquist - 2.7 kHz. See it? Returning to Audition so that I can get better resolution, I find that as close as I can get to 19.35 (Nyquist - 2.7kHz), the level is -73.85. What I'm getting at is that, if you start with the assumption of a symmetric spectrum--aliasing--then the peak at Nyquist + 2.7kHz is down by about 16dB relative to Nyquist - 2.7kHz. Note also that with any of the filters in question, 19.35kHz--Nyquist - 2.7--should be essentially flat.
Now let's go to JA's measurements. Look at the MQA filter. Resolution is poor (because I'm reading a small graph and the line has significant spread), but to me it looks like "flat" on the filter is at about -43dB. At about 25kHz, the filter is down to about -59dB by my estimate; it really is hard to tell. Take a look and see if you agree. That would mean that the filter is down about 16dB at Nyquist + 2.7kHz, which is exactly what we observed above.
This spectrum matches up well with the MQA filter.
Now--finally--take a look at the image below. In the Beyonce spectrum (that's CD and MQA versions; you can barely see the red shining through the blue), I've zeroed in on the region near the Nyquist frequency. Look at it closely. Start in the middle and work your way out toward the sides. Except for the fact that it's increasingly attenuated at higher frequencies, the symmetry is, I think, apparent.
That is aliasing.
> > That is aliasing. < <
Yep, agree 100%. Thanks for your patience and diligence with this. And I also see now why I said "something is very wrong here". Your explanation detailed in your post is right on the money. What is wrong about the graph is that it seems quite apparent that MQA sent a single-rate signal through a dual-rate filter.
This is essentially the equivalent of using a so-called "NOS" (non-oversampling) D/A converter. The only "reconstruction filter" in those cases is simply due to the bandwidth limitations of the analog circuitry. (In some cases the analog bandwidth for NOS DACs is deliberately restricted - say by using transformers in the audio signal path.)
Now all of these graphs make sense, and I would assert that there *is* a bug in this implementation of the MQA system. Specifically when a dual- or quad-rate source file is used, the digital filter used is a slow rolloff type that may allow for some *slight* "leakage" which contributes to very low levels of aliasing. But when a single-rate source file is used, the dual-rate MQA filter provides so little filtering as to act more like a filterless ("NOS") D/A converter. All of the energy in the top octave is simply non-harmonic (ie, unpleasant sounding) distortion. It doesn't make sense to me that one would sharply reduce aliasing at clearly inaudible frequencies above 48kHz, yet allow high levels of aliasing in potentially more troublesome frequencies an octave lower.
Furthermore we can now understand why some single-rate MQA files exhibit the notch centered at 22.05kHz. If the original source file was sampled at 44.1kHz, subsequent MQA processing apparently occurs at 48kHz and all content above 22.05kHz (=Fs/2, the Nyquist frequency) is simply aliasing that is only mildly attenuated. The same is true if the original source file was sampled at 48kHz, except there would be no gap between the original cutoff due to the anti-aliasing filter in the A/D converter and the mirror-imaged aliasing (purely artifacts comprising non-harmonic distortion) would connect seamlessly with the original audio.
In this case if a listener prefers the sound of an MQA-processed single-rate signal to that of the original, I assert they would likely prefer the sound of a filterless ("NOS") D/A converter even more. Apparently the main difference created by the MQA processing of a single-rate source file is the addition of the non-harmonic distortion created by the aliasing.
When I spent many months auditioning digital filters roughly a decade ago, one of the first things I tested was a filterless ("NOS") solution, as at that time they were fairly popular and making a "buzz" in the market. While there were many attractive aspects about the sound quality, in my opinion the filter we settled on retained the good qualities of the filterless approach while improving significantly on what I felt were its weaknesses.
As always, strictly my own opinions and not necessarily those of my employer or favorite athlete.
In an older post on the "Digital" forum you had written "MQA claims an end-to-end (which I take to mean encompassing the ADC and the DAC)". I just now ran across some information that clarifies what MQA means by "end-to-end". The following is from a FAQ on the Auralic website (full link below):
"AURALiC has done a live demo during CES 2016 for MQA on ARIES and ARIES MINI. It is however after MQA realized that ARIES does not have any DAC built-in and ARIES MINI has a digital output in parallel connection of its DAC I2S signal, they pulled it back immediately. They believe the MQA process is end to end and the DAC has to be optimized for MQA playback, so any digital output of fully decoded signal is unacceptable ." [Emphasis added.]
One interpretation of this is that MQA believes it is critically important to "compensate" for the specific characteristics of the DAC chip used in a given A/D converter. A side affect of this policy is that it forces MQA to be a closed system. Without this restriction it would be trivially easy to create outboard MQA decoders (either hardware or software) and use them with any existing D/A converter - obviating the need to purchase an entirely new D/A converter.
This restriction also seems to support the view of MQA as a form of DRM - the SACD format also specifically disallowed the digital signal to be accessed externally. (This was one of Sony's main selling points of SACD to the record labels.) Fully decoded MQA files are only available to those who purchase MQA-licensed hardware, which includes a royalty payment to MQA.
Even if MQA felt it critically important to compensate for the DAC chip, it would seem they could instead choose to use an outboard decoder along with a some other means to optimize for a specific DAC chip. One choice would be a model similar to what Devialet provides for loudspeaker DSP correction. One can simply select from one of (currently) 736 supported models of loudspeakers and receive the "proper correction" via Devialet's SAM feature (Speaker Active Matching):
As always, strictly my own opinions and not necessarily those of my employer or city mayor.
Charles, I lack your experience-based perspective, so I guess to me it doesn't seem that odd. I go back to (one of) the patent application(s) and see this language:
> > Preferrably, the downsampler comprises decimation filter specified at the first sample rate, wherein the asymmetric component of the response of the decimation filter is characterized by an attenuation of at least 32dB at frequencies that would alias to the 0-7kHz range.
The range 0-7kHz is where the ear is most sensitive. The amount of attenuation required varies greatly according to the spectrum to be encoded in the vicinity of its Nyquist frequency, and may [sic] signals will require more than 32dB of attenuation. < <
For CD-res, 7kHz is Nyquist - 15kHz. So what's the attenuation of that MQA filter JA measured at Nyquist + 15kHz = 37.1kHz? It looks to be down about 55dB, give or take, which even with this slow roll-off easily meets that spec. Looking at the Beyonce track--comparing 7kHz with 37kHz--just eyeballing it this time, not going back to the Audition file--the levels seem to be about -70dB at 7kHz and about -120dB at 37kHz--so, again, aliasing spec easily met--even with this very gentle filter.
MQA has been quite explicit about the desire to trade aliasing for time resolution. They appear to be doing just what they said they would do--so I don't see why you'd call it a bug. Surely they'd call it a feature.
One thing I don't understand is why you say they're using a double-rate filter. The filter shown in JA's Fig. 6 accurately describes the data, is flat not quite to the 44.1 Nyquist frequency (that is, about 22kHz). It describes what's happening but it would never work for 96kHz data.
Otherwise, we seem to have converged on an interpretation.
OK, now I get it!
I was wrong and you were right, and now I understand where I was led astray with regards to the aliasing. I only read JA's Brooklyn review close enough to notice the single-rate MQA filter response just recently (even though it is ~9 months old) and it hadn't fully sunk in. Instead I've been operating under the (presumably still accurate) premise that the limiting factor in the MQA "end-to-end system response" was the digital reconstruction filter (in the D/A converter), which at the quad-rate sampling frequency yields this frequency response curve:
and mistakenly assuming that was the filter also being used with single-rate audio data. I now completely agree with your analysis using your measured spectra combined with (the other) JA's measurements of the Brooklyn single-rate filter.
And now we also an see the likely reason why the notch appears in the spectrum of single-rate MQA files originally recorded at 44.1kHz. Virtually all digital equipment (recording or playback) uses "halfband" digital filters, as these require only half as many taps and half as much storage for the memory coefficients. One key characteristic of a halfband filter is that it will always have a response of -6dB at the Nyquist frequency (Fs/2).
If that output is later fed to a "leaky" reconstruction filter that mirror-images the audio data as out-of band aliases, there will be a continuous graph with no gap between the original music and the aliased version. But that is not what we see in your Beyonce track. Instead it appears that at some point the audio data was fed through an "apodizing" filter, specifically designed to filter out the "ringing" created by the anti-aliasing filter of the A/D converter at the Nyquist frequency (Fs/2). I would guess that this was part of the MQA process, as I am unaware of any commercially available A/D converters with built-in "apodizing" filters.
When that filtered audio data is then sent through yet another digital filter - this time the "leaky" MQA filter in the D/A converter, the sharp rolloff between 21kHz and 22kHz seen in your original spectrum of the Beyonce track is mirror-imaged at 22.05kHz due to the aliasing. The combination of the "apodizing" filter and the "leaky" MQA filter creates the "notch" seen in the spectrum. Thank you for your input in helping to solve this mystery!
As far as MQA's claim that 32dB of anti-aliasing is sufficient for signals in the 0 (DC) to 7kHz range, I think that is open for debate. It would seem that high-performance audio generally revolves around the idea of continuous improvement, and not dictating "sufficient" levels of performance.
As always, solely my own opinions and not necessarily those of my employer or butcher.
Charles, thanks for this. I may have been right (approximately) from early on, but exchanges like this help solidify my tenuous knowledge and provisional thinking. And it should be noted that any insight I have derives also from conversations with--and in some cases specific ideas expressed by--the other JA.
Your explanation here is consistent with one thing Bob Stuart has said repeatedly--that what MQA does depends on the recording. The other 44.1 MQA I posted results from didn't have this wide, obvious gap. Presumably the Beyonce needed a lot of apodizing. Curious, then, that the CD version and the MQA version should be nearly identical--what do you make of that? (The possibility that this near-identity is an implementation error--a mistake or compromise made by this particular converter and not necessarily a specifically MQA phenomenon--of course remains.)
Just one more comment. You wrote:
As far as MQA's claim that 32dB of anti-aliasing is sufficient for signals in the 0 (DC) to 7kHz range, I think that is open for debate. It would seem that high-performance audio generally revolves around the idea of continuous improvement, and not dictating "sufficient" levels of performance. < <
But the MQA folks maintain--this is my interpretation; I hope it's more or less correct--that this is not "settling", but, rather, striking a different compromise, one that prioritizes the time domain over the frequency domain to a greater extent than any previous technology. It could therefore be seen as a natural extension of the last couple of decades of thinking about digital audio, by you and others. I remember reading, a few years ago, Ayre's White Paper on minimum phase filters. Viewed in that context, MQA could be seen as a natural next step--although to say it's natural is not necessarily to say it's wise. Plus, I've seen no direct evidence that they're achieving the time-domain performance they've specified. It's on this basis--that the technology is plausible and a reasonable (if radical) next step, that I've argued that it deserves an audition.
> > exchanges like this help solidify my tenuous knowledge and provisional thinking < <
Same here. One would think that MQA would simply explain the technical details clearly for all to understand, but I am also trying to "solidify my tenuous knowledge and provisional thinking".
> > Curious, then, that the CD version and the MQA version should be nearly identical--what do you make of that? < <
Again I would like more data to avoid jumping to the incorrect conclusion (as I did previously). Is the "CD version" actually a physical disc? Whatever its source, it has clearly been run through an apodizing filter, directly in contradiction to Peter Craven's (originator of the apodizing filter) recommendations (see below).
It seems unlikely that there is some obscure "pro" manufacturer making a "state-of-the-art" A/D converter with an apodizing anti-aliasing filter mistakenly attempting to wring the very best sound out of 44.1kHz sample rates. To the best of my knowledge the first proposed use of an apodizing filter was in Peter Craven's AES paper of 2004. I believe the first commercial use was likely by Meridian a few years later and know that Ayre experimented with them during 2008 and released them as the new "Listen" filter in the MP upgrade of 2009 (and also the QB-9 and subsequent digital products). But all of these were reconstruction filters for use on the D/A side.
I would think the most likely possibility is that MQA had something to do with the application of the apodizing filter to this Beyonce release. Which raises at least three questions in my mind:
1) If MQA used an apodizing filter to remove any pre-ringing from the A/D converters used to create the Beyonce CD, why not simply sell, license, use, or give away that technology to improve the sound of all recordings? Or is there some other company or mastering engineer that is already doing the same thing but without using MQA's tools?
2) It would seem that the main change upon playback between the Beyonce CD and the MQA-encoded version of the same file is that upon playback, the "leaky" MQA filter is used. In your original trace of the CD version, I saw no evidence of aliasing - presumably because the reconstruction filter was not "leaky". In contrast the "leaky" MQA filter passed higher levels of aliasing artifacts than apparently even the "Slow" filter on the Mytek Brooklyn. (Again, I am unclear on all of the details of the test conditions.) But if that is the case and the Beyonce CD already has the "time blur" filtered out in the non-MQA version, all we have learned is that different digital reconstruction filters sound different - which wouldn't seem to be a revolutionary breakthrough.
3) If an apodizing filter has been applied to the Beyonce CD, this is in direct contradiction to Peter Craven's recommendation in his AES paper. Specifically, any filter sharp enough to filter out the "pre-ringing" introduced by the A/D converter will also introduce "ringing" of its own. The only advantage of the apodizing filter is that being minimum-phase, all of its "ringing" will be more natural sounding "post-ringing" rather than the nowhere-to-be-found-in-nature "pre-ringing" created by linear-phase filters.
Again, apparently More Questions than Answers...
As always, strictly my own opinions and not necessarily those of my employer or pet wombat.
Busy today, unfortunately. I'd love to pursue this further. I won't get to that until tomorrow probably.
I can answer one question, since it takes no time: All the measurements I presented are from Tidal streams. There are two versions of "Lemonade" on Tidal (Hi-Fi/Master): the MQA version and a regular CD-res version. That's what I've presented, delivered by Roon, apparently bit-perfect (since the Blue MQA lights up on the MQA file).
I appreciate all of your work in putting this information together. Definitely an interesting puzzle. The thing that has me scratching my head is the application of an apodizing digital filter to the "CD-res version" of the Beyonce album. This is a no-no, according to Peter Craven's original paper describing apodizing filters:
"We suggest that the final digital to analogue conversion may be
the appropriate place for the apodising filter."
That is from the summary conclusion. Earlier in the paper is an entire section explaining the reasons for the end of the chain as the optimal location for the apodizing filter.
As always, strictly my own opinions and not necessarily those of my employer or minister.
Post a Message!
This post is made possible by the generous support of people like you and our sponsors: